ProgrammableMarketing/mangodesk.com (opens in a new tab)
PDF

Views

  • Plan
  • Roadmap
  • Research

On this page

  • The brief
  • Start here
  • Idea portfolio
  • Effort vs impact
  • This week · 6
  • This quarter · 11
  • Big bets · 3
  • Wildcards · 2
  • How we can help
Generated Nov 10, 2026
  1. Programmable
  2. Playbooks
  3. MangoDesk
MangoDesk logo

growth playbook · Nov 10, 2026

mangodesk.com (opens in a new tab)

MangoDesk growth playbook

Diagnosis

MangoDesk is a two-person YC Summer 2025 company with strong founder credentials (ex-Scale AI, a previous seven-figure software business) selling into one of AI's hottest budgets: RL environments for post-training. Its public presence hasn't kept up. The site is a single Lovable page with an email link as the only route to buy.

Its privacy, terms, trust, pricing and changelog URLs all return 404. The mobile page takes 30 seconds to paint its headline because of 5.7 MB of images. The company describes itself four different ways across its site, YC page, launch post and PyPI package. And it is missing from RL List, the 42-vendor directory lab buyers use, while rivals such as Proximal, Halluminate and Datacurve lead with public benchmarks.

This playbook first fixes the buying path (a buyer page, legal and trust pages, a scoping form, one positioning line, faster images, working package links). Next it builds proof: a public benchmark, quality reports, a fixed-scope pilot and environment pages. Later bets are a vertical focus, self-serve pipelines and marketplace distribution.

Growth Score

4/10

Biggest Opportunity

Lab buyers are actively shortlisting RL-environment vendors and score them on named customers, security posture and primary-source proof. A credible buyer page, trust pack and public benchmark would let MangoDesk's founder pedigree show up where those shortlists get made.

Biggest Risk

The category is crowded: 42 vendors on RL List, plus incumbents like Scale AI, Mercor, Surge and Turing. Proximal, Halluminate and Datacurve already publish benchmarks. With a two-person team, MangoDesk could stay invisible to buyers who never get past the one-page site and missing legal pages.

01

Start here

The three moves to make first. Each one is written up in full below.

  1. 1

    Replace the one-line homepage with a buyer page, and put live privacy, terms and trust pages where today's links return 404

    Publish privacy and terms pages at /privacy and /terms today, then write the three sample-environment blurbs.

    High impact First result in 2-4 weeks
  2. 2

    Replace the founders' email link with a four-question scoping form for RL-environment projects

    Write the four questions and set up a simple form with email routing.

    High impact First result in 2 weeks
  3. 3

    Get onto RL List, the 42-vendor directory lab buyers use, with primary-source facts

    Write the vendor profile with a mangodesk.com source link for each field.

    High impact First result in 1-2 months

02

Your idea portfolio

22 ideas, grouped by when to run them. Open any idea for how to run it: the steps, who to approach, copy you can use, and how to measure it.

  • 7 ideas are marked Unexpected: plays you are unlikely to have tried.

    1 of these are things Programmable can build and run for you.

Every idea placed by the effort it takes against the impact it could have. Quick wins sit top left; select a number to open that idea. The filters below narrow the list by status, timing, category and what Programmable can build.

Effort vs impactEach number is an idea. Select one to open it.
Impact →
Quick wins · 4Big bets · 7Fill-ins · 6Reconsider · 5
Effort →

Showing 22 of 22 ideas

This week 6

Quick wins with a clear first action.

  • 01

    A lab buyer arriving at mangodesk.com gets one sentence ("We evaluate and improve AI on meaningful use cases via production grade RL environments"), a YC logo and a mailto link. There's nothing about which workflows MangoDesk covers, what a delivered environment contains, how quality is verified, or who has used it. /privacy, /terms, /trust, /pricing and /changelog all return the site's 404 page, and there is no sitemap. Procurement at an AI lab checks for privacy and data-handling terms before anything else. RL List also scores vendors on named customers, security posture and primary-source facts, and this site gives it none of those to cite.

    High impactLow effort First result in 2-4 weeks Co founder & CEO

    First action

    Publish privacy and terms pages at /privacy and /terms today, then write the three sample-environment blurbs.

    How to measure it

    Inbound scoping calls from mangodesk.com

    Now
    –
    Target
    4 a month
    By
    60 days

    Steps

    1. 1

      Fix the 404s

      Publish /privacy and /terms; remove or fill /pricing, /trust and /changelog.

      Co-founder & CTO 1 day

    2. 2

      Write the trust page

      Data handling, access, hosting, certification status and a security contact.

      Co-founder & CTO 1 day

    3. 3

      Pick three sample environments

      One task, verifier and success criterion for each, scrubbed of customer data.

      Strategic Projects Lead 2 days

    4. 4

      Write the buyer page

      Positioning, environment types, engagement steps, quality method, team.

      Co-founder & CEO 2 days

    5. 5

      Add a sitemap

      sitemap.xml listing every live page; submit it in Search Console.

      Founding Engineer 2 hours

    6. 6

      Link the scoping form

      Replace the mailto with the form from idea 6.

      Founding Engineer 2 hours

    7. 7

      Measure

      Track visits to form starts and to calls.

      Co-founder & CEO monthly

    First result
    2-4 weeks
    Cost
    Founder time; a privacy and terms template or a short legal review
    Owner
    Co founder & CEO

    Smallest test

    Publish the privacy and terms pages and one sample environment this week, and send the page to the next three prospects.

    Why it works

    Rebuild the homepage as a buyer page in one week, on the existing Lovable site. Sections: (1) one positioning line (see idea 3); (2) three environment types with one sample task each, e.g. a long-horizon software-engineering task with its verifier; (3) how an engagement works: scope, build, verify, deliver, refresh; (4) quality: how rewards are verified and checked for hacking; (5) the team's background and YC; (6) a scoping form (idea 6). Publish real /privacy and /terms pages, plus a /trust page covering data handling, access controls, where environments run and the current certification status, even if that is "SOC 2 in progress". Add a sitemap.xml. Remove or fill any URL that currently 404s.

    Lab buyers shortlist vendors before they ever email. A page that shows a concrete sample environment and clears the legal check turns a curious visitor into a scoped conversation, and the same pages give directories and AI assistants primary-source facts to cite.

    Ready to use

    ChecklistBuyer page launch list

    [ ] /privacy live (not 404)
    [ ] /terms live
    [ ] /trust: data handling, access, hosting, certification status, security contact
    [ ] /pricing removed or replaced with the pilot offer (idea 9)
    [ ] /changelog removed or filled (idea 17)
    [ ] One positioning line in title, meta and H1
    [ ] Three sample environments with task, verifier, success criterion
    [ ] Engagement steps: scope, build, verify, deliver, refresh
    [ ] Scoping form replaces mailto
    [ ] sitemap.xml submitted

    Risks and how to handle them

    • Sample environments reveal customer work

      Use synthetic tasks built only for the site.

    • The trust page over-promises

      State only current controls and label anything in progress.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
    • Competitor researchRivals' own homepages: Proximal leads with its FrontierSWE v2 benchmark ("34 tasks, 13 frontier models, 20-hour budgets"), a public leaderboard, a blog and Privacy and MSA pages; Halluminate leads with an 88-problem finance due-diligence benchmark "written and reviewed by practicing finance deal professionals" and has Research, Expert Network and Data Partnerships pages; Datacurve promotes its "DeepSWE" long-horizon coding benchmark and a research page; Fleet links Privacy, Terms, Cookies and a status page.Source
  • 02

    In Google's mobile test the homepage's main content (the "Measurable model improvement" block) takes 30.5 seconds to appear. The page weighs 5,957 KB, and 5,738 KB of that is images on a page with only 10 requests. PageSpeed estimates 5,222 KB of savings from better image delivery. Desktop is fine at 2.6 s, and there's no real-visitor data yet, so this is a lab result, but it's the one an investor or lab researcher opening a link on their phone will feel.

    Medium impactLow effort First result in 1 day Founding Engineer

    First action

    List every image on the homepage with its file size and display size.

    How to measure it

    Mobile LCP (PageSpeed lab)

    Now
    –
    Target
    under 2.5 s
    By
    2 weeks

    Steps

    1. 1

      Inventory images

      File size against display size for each.

      Founding Engineer 1 hour

    2. 2

      Re-export

      WebP or AVIF at 1x and 2x display size.

      Founding Engineer 2 hours

    3. 3

      Lazy-load

      Below-the-fold images only.

      Founding Engineer 1 hour

    4. 4

      Set caching

      Long cache lifetimes for static assets in Cloudflare.

      Founding Engineer 1 hour

    5. 5

      Re-test

      PageSpeed mobile and desktop.

      Founding Engineer 15 minutes

    6. 6

      Guard it

      A size limit for future uploads.

      Co-founder & CTO 30 minutes

    First result
    1 day
    Cost
    $0
    Owner
    Founding Engineer

    Smallest test

    Replace the single largest image and re-run PageSpeed.

    Why it works

    Convert the images under /lovable-uploads/ to WebP or AVIF at the size they're actually displayed (the logo and the YC logo shouldn't be megabytes). Lazy-load anything below the fold and set long cache lifetimes on static assets in Cloudflare. Re-run PageSpeed and aim for mobile LCP under 2.5 s.

    It's a one-day fix. A first impression that loads instantly matters for a company selling "production-grade" infrastructure.

    Risks and how to handle them

    • Lovable regenerates assets on the next edit

      Keep optimized images in the repo and check after each publish.

    Evidence

    • PageSpeedGoogle PageSpeed, November 10, 2026: mobile performance 72 (desktop 84); Largest Contentful Paint 30.5 s on mobile and 2.6 s on desktop, the LCP element being the "Measurable model improvement" text block; the page weighs 5,957 KB, 5,738 KB of it images, across 10 requests; "Improve image delivery" could save 5,222 KB and longer cache lifetimes 2,821 KB. Accessibility, best practices and SEO all score 100. There is no real-visitor data yet, so these are lab results.Source
    • Your websiteSite build (programmable.info scan, November 10, 2026): the homepage is built with Lovable (meta author "Lovable", images under /lovable-uploads/) and hosted on Render behind Cloudflare; no analytics, blog, newsletter, live chat or social profile links detected.Source
  • 03

    MangoDesk currently describes itself four ways. The page title says "Accelerate custom evals and data". The meta description says "evals and data pipelines with natural language". The H1 says "production grade RL environments", and YC says "Long-horizon RL environments" for "real-world workflows". The launch listing calls it "Cursor for Evals and Human Data" and the PyPI package "tau-bench wrapper with cloud dashboard". Search results, link previews and AI assistants pick up whichever line they find first, so buyers get mixed signals about what MangoDesk sells.

    Medium impactLow effort First result in 2 weeks Co founder & CEO

    First action

    Write three candidate lines and test them on two current customers or prospects.

    How to measure it

    Profiles using the agreed line

    Now
    –
    Target
    6 of 6
    By
    2 weeks

    Steps

    1. 1

      List every description

      Site title, meta, H1, YC, LinkedIn, PyPI, launch post.

      Co-founder & CEO 1 hour

    2. 2

      Draft lines

      Three options built around RL environments.

      Co-founder & CEO 1 hour

    3. 3

      Test

      Ask two buyers which one describes what they bought.

      Strategic Projects Lead 2 days

    4. 4

      Update the site

      Title, meta, H1, og tags.

      Founding Engineer 1 hour

    5. 5

      Update profiles

      YC, LinkedIn, PyPI summary on the next release.

      Co-founder & CEO 1 hour

    6. 6

      Check previews

      Link previews and search snippets.

      Founding Engineer 1 week

    First result
    2 weeks
    Cost
    $0
    Owner
    Co founder & CEO

    Smallest test

    Change the title and meta description to match the H1 today.

    Why it works

    Choose one line that matches what is sold today (for example, "Long-horizon RL environments and evals for real-world software and knowledge work") and use it as the page title, meta description, H1, YC one-liner, LinkedIn tagline and PyPI summary. Keep "evals" and "post-training data" as supporting phrases, since buyers search for those words too.

    One consistent line is what directories, journalists and AI summaries repeat, and it helps buyers recognise MangoDesk when they compare vendors.

    Risks and how to handle them

    • Narrowing loses the evals buyers

      Keep evals in the supporting sentence and the meta description.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • Y CombinatorYC company page: Summer 2025 batch, team size 2, San Francisco; one-liner "Long-horizon RL environments"; "We help AI become better at real-world workflows through production-grade RL environments."; launch post titled "MangoDesk: Cursor for Evals and Human Data"; links a LinkedIn page and no X account.Source
    • PressLaunch listing (August 13, 2025): "MangoDesk: Cursor for Evals and Human Data", to "Accelerate creating custom evals and post-training data by spinning up bespoke data-annotation pipelines in seconds"; company line "Simple platform to create data for AI."Source
    • PyPIPackage "mangodesk" 0.5.4: "tau-bench wrapper with cloud dashboard and trajectory tracking" that runs evaluations in retail, airline and telecom environments and uploads results to MangoDesk's cloud; proprietary licence; nine releases between December 30, 2025 and January 2, 2026, none since. Its Documentation link (docs.mangodesk.com) does not resolve and its GitHub Repository link returns 404 (checked November 10, 2026).Source
  • 04

    RL List ranks and tracks 42 RL-environment vendors for frontier labs, with category pages for coding, computer use and enterprise. It lists direct rivals such as Proximal, Halluminate and Datacurve, but not MangoDesk. Its score rewards named or verified customers, frontier-lab ties, security posture such as SOC 2, and a bonus for every field backed by a primary source. Today MangoDesk's own site provides none of those sources.

    High impactLow effort First result in 1-2 months Co founder & CEO

    First action

    Write the vendor profile with a mangodesk.com source link for each field.

    How to measure it

    Vendor directories listing MangoDesk

    Now
    –
    Target
    3
    By
    60 days

    Steps

    1. 1

      Wait for sources

      Buyer page and trust page live first (idea 1).

      Co-founder & CEO 1 week

    2. 2

      Draft the profile

      Each field linked to a mangodesk.com page.

      Co-founder & CEO 2 hours

    3. 3

      Pick categories

      Coding and enterprise environments.

      Strategic Projects Lead 1 hour

    4. 4

      Submit

      Through the directory's public contact route.

      Co-founder & CEO 1 hour

    5. 5

      Find other lists

      YC industry page, analyst and newsletter vendor maps.

      Strategic Projects Lead 1 day

    6. 6

      Keep current

      Update each listing when a benchmark, certification or customer is added.

      Co-founder & CEO quarterly

    First result
    1-2 months
    Cost
    $0
    Owner
    Co founder & CEO

    Smallest test

    Send the sourced profile once the trust page is live.

    Why it works

    Once the buyer and trust pages (idea 1) are live, prepare a one-page vendor profile: what you build, the environment categories you cover (software engineering, knowledge work), team background, funding, public benchmark or sample (idea 7) and security status. Each fact should link to a page on mangodesk.com. Contact the directory through its public channels and ask to be added to the coding and enterprise categories. Repeat for the YC Reinforcement Learning industry page and other public vendor lists.

    Buyers and analysts build shortlists from directories like this, and the score explicitly rewards the primary-source proof MangoDesk can publish this month.

    Risks and how to handle them

    • A low initial rank

      Add primary sources over time; the score is explicitly not a quality rating.

    Builds on

    Evidence

    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • Y CombinatorYC company page: Summer 2025 batch, team size 2, San Francisco; one-liner "Long-horizon RL environments"; "We help AI become better at real-world workflows through production-grade RL environments."; launch post titled "MangoDesk: Cursor for Evals and Human Data"; links a LinkedIn page and no X account.Source
  • 05

    The mangodesk package on PyPI (0.5.4) wraps tau-bench with a cloud dashboard and trajectory tracking, and it's probably the first thing a technical evaluator at a lab installs. Its Documentation link (docs.mangodesk.com) doesn't resolve, its Repository link on GitHub returns 404, and there have been no releases since January 2, 2026. An engineer who tries it hits a dead end and may conclude the product is abandoned.

    Medium impactLow effort First result in 1 week Founding Engineer

    First action

    Decide whether the package is still part of the offer.

    How to measure it

    Visits from PyPI and docs links

    Now
    –
    Target
    tracked and growing
    By
    60 days

    Steps

    1. 1

      Decide

      Revive as a free entry point, or retire.

      Co-founder & CTO 1 day

    2. 2

      Write docs

      Install, run, view results, on mangodesk.com/docs.

      Founding Engineer 1 day

    3. 3

      Fix links

      Docs and repository URLs in the package metadata.

      Founding Engineer 1 hour

    4. 4

      Release

      A patch version with the new links and a changelog entry.

      Founding Engineer 2 hours

    5. 5

      Tag traffic

      UTM links from the README to the site.

      Founding Engineer 30 minutes

    6. 6

      Review

      Downloads and dashboard sign-ups each month.

      Co-founder & CTO monthly

    First result
    1 week
    Cost
    $0
    Owner
    Founding Engineer

    Smallest test

    Release a version whose links point to working pages.

    Why it works

    Either revive the package or retire it cleanly. To revive it: publish a short docs page on mangodesk.com (install, run an eval, view results in the dashboard), point the PyPI links at pages that exist, make the repository public or remove the link, and ship a release noting current tau-bench compatibility. To retire it: publish a final release whose README says it's no longer maintained and points to the buyer page.

    Developers judge a vendor by whether its tools work. A working docs link turns installs into conversations instead of bounces.

    Risks and how to handle them

    • Maintenance load on a two-person team

      Retire it cleanly if it isn't part of the offer.

    Evidence

    • PyPIPackage "mangodesk" 0.5.4: "tau-bench wrapper with cloud dashboard and trajectory tracking" that runs evaluations in retail, airline and telecom environments and uploads results to MangoDesk's cloud; proprietary licence; nine releases between December 30, 2025 and January 2, 2026, none since. Its Documentation link (docs.mangodesk.com) does not resolve and its GitHub Repository link returns 404 (checked November 10, 2026).Source
  • 06

    The only way to buy from mangodesk.com is "email [email protected]". A blank email asks a busy post-training lead to write the brief themselves, and the founders get no structured information to prioritise with. MangoDesk is now hiring a Strategic Projects Lead to own delivery for strategic customers, which is the natural owner for a scoping intake.

    High impactLow effort First result in 2 weeks Strategic Projects Lead

    First action

    Write the four questions and set up a simple form with email routing.

    How to measure it

    Completed scoping forms

    Now
    –
    Target
    6 a month
    By
    90 days

    Steps

    1. 1

      Write questions

      Four, all answerable in under two minutes.

      Strategic Projects Lead 2 hours

    2. 2

      Build the form

      On the homepage and the trust page.

      Founding Engineer 2 hours

    3. 3

      Route

      To the Strategic Projects Lead with a founder copy.

      Founding Engineer 1 hour

    4. 4

      Scope template

      A one-page reply: tasks, verifiers, timeline, price range.

      Strategic Projects Lead 1 day

    5. 5

      Log requests

      Domain, type, stage and outcome.

      Strategic Projects Lead ongoing

    6. 6

      Review

      Top requested domains each month.

      Co-founder & CEO monthly

    First result
    2 weeks
    Cost
    $0
    Owner
    Strategic Projects Lead

    Smallest test

    Add the form alongside the email link for a month and compare.

    Why it works

    Add a short form: (1) the workflow or domain to train or evaluate on, (2) environments or evals, (3) the model stage (pre-release, post-training, production agent), (4) timeline and rough volume. Route answers to the Strategic Projects Lead, reply within one business day with a one-page scope, and log every request so recurring domains become standard environment packages (idea 10).

    Structured inbound converts better than a blank email. It also shows which domains buyers ask for, which feeds the catalogue and pricing.

    Risks and how to handle them

    • Labs won't put details in a form

      Keep the email link as an alternative and keep questions high level.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • HiringStrategic Projects Lead (San Francisco, on-site, posted August 14, 2026): leads cross-functional projects across engineering, operations and go-to-market, owns delivery of RL environments for strategic customers and reports directly to the founders; needs a technical background (at least SQL or Python data analysis), with human data or RL experience preferred.Source

This quarter 11

The core plays for the next three months.

  • 07

    MangoDesk's closest rivals lead with public benchmarks: Proximal with FrontierSWE v2 ("34 tasks, 13 frontier models, 20-hour budgets") and a leaderboard, Halluminate with an 88-problem expert-written finance benchmark, and Datacurve with DeepSWE. A press feature on MangoDesk said "the actual quality of the generated data is the only metric that matters" and named no customers. A public benchmark is the proof a lab can inspect without signing an NDA.

    High impactHigh effort First result in 2-3 months Co founder & CTO

    First action

    Choose the domain and write ten tasks with verifiers.

    How to measure it

    Held-out evaluation requests

    Now
    –
    Target
    3
    By
    2 quarters

    Steps

    1. 1

      Pick the domain

      A workflow rivals don't benchmark.

      Co-founder & CEO 1 week

    2. 2

      Write tasks

      20-40 long-horizon tasks with verifiers.

      Strategic Projects Lead 4 weeks

    3. 3

      Harden verifiers

      Check for reward hacking and ambiguity.

      Founding Engineer 2 weeks

    4. 4

      Run models

      Current public models under fixed budgets.

      Founding Engineer 1 week

    5. 5

      Publish

      Method, examples, leaderboard on mangodesk.com.

      Co-founder & CTO 1 week

    6. 6

      Promote

      Post the results; send them to lab researchers and directories.

      Co-founder & CEO 2 weeks

    7. 7

      Refresh

      New models and tasks each quarter.

      Founding Engineer quarterly

    First result
    2-3 months
    Cost
    Compute for model runs plus several weeks of team time
    Owner
    Co founder & CTO

    Smallest test

    Publish ten tasks and results for three models.

    Why it works

    Build a small benchmark (20-40 tasks) in the area MangoDesk wants to own: long-horizon knowledge-work or software workflows that rivals don't cover. Run current public models on it, publish the method, verifiers and a leaderboard on mangodesk.com, and offer labs private evaluation on a held-out set. Refresh it every quarter.

    Benchmarks are how this market judges quality, and every result a lab cites is free distribution with MangoDesk's name on it.

    Risks and how to handle them

    • Tasks leak into training data

      Keep a held-out set private.

    • Distracts from delivery

      Reuse tasks from existing environment work where rights allow.

    Evidence

    • Competitor researchRivals' own homepages: Proximal leads with its FrontierSWE v2 benchmark ("34 tasks, 13 frontier models, 20-hour budgets"), a public leaderboard, a blog and Privacy and MSA pages; Halluminate leads with an 88-problem finance due-diligence benchmark "written and reviewed by practicing finance deal professionals" and has Research, Expert Network and Data Partnerships pages; Datacurve promotes its "DeepSWE" long-horizon coding benchmark and a research page; Fleet links Privacy, Terms, Cookies and a status page.Source
    • PressHuge Magazine feature (February 2026): names no MangoDesk customers or design partners; says generation speed does not prove output quality and "the actual quality of the generated data is the only metric that matters"; plans to watch generated-pipeline quality against hand-built alternatives, domain edge cases and retention.Source
  • 08

    The main doubt about generated environments and data is quality. The February feature said speed doesn't prove it and listed pipeline quality against hand-built alternatives as the thing to watch. MangoDesk's homepage promises "measurable model improvement" but shows no measurements. The new Strategic Projects Lead owns delivery for strategic customers, which is where a standard quality report would live.

    High impactMedium effort First result in 1 month Strategic Projects Lead

    First action

    Draft the report template on the most recent delivery.

    How to measure it

    Share of deliveries with the report

    Now
    –
    Target
    100%
    By
    1 quarter

    Steps

    1. 1

      Define measures

      Agreement, solvability, hacking probes, difficulty.

      Strategic Projects Lead 1 week

    2. 2

      Automate

      Scripts that generate the numbers.

      Founding Engineer 2 weeks

    3. 3

      Pilot

      On the next delivery.

      Strategic Projects Lead 2 weeks

    4. 4

      Ask for training results

      Before/after scores where the customer agrees.

      Strategic Projects Lead ongoing

    5. 5

      Publish a sample

      An anonymised report on mangodesk.com.

      Co-founder & CEO 1 week

    6. 6

      Use in sales

      Attach to every scope.

      Co-founder & CEO ongoing

    First result
    1 month
    Cost
    Engineering time
    Owner
    Strategic Projects Lead

    Smallest test

    Produce the report by hand for one delivery.

    Why it works

    Define a one-page report delivered with each environment: task count and difficulty spread, verifier agreement with expert graders on a sample, solvability (a reference solution passes), known reward-hacking probes and results, and, where the customer allows, model scores before and after training. Publish an anonymised sample report on the site as proof.

    It turns MangoDesk's main objection into a selling point, and it gives buyers a standard to compare other vendors against.

    Risks and how to handle them

    • Numbers look weak at first

      Share internally first and improve before publishing.

    Evidence

    • PressHuge Magazine feature (February 2026): names no MangoDesk customers or design partners; says generation speed does not prove output quality and "the actual quality of the generated data is the only metric that matters"; plans to watch generated-pipeline quality against hand-built alternatives, domain edge cases and retention.Source
    • HiringStrategic Projects Lead (San Francisco, on-site, posted August 14, 2026): leads cross-functional projects across engineering, operations and go-to-market, owns delivery of RL environments for strategic customers and reports directly to the founders; needs a technical background (at least SQL or Python data analysis), with human data or RL experience preferred.Source
    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
  • 09

    mangodesk.com/pricing returns 404, and the only commercial route is a custom conversation with the founders. For a lab trying a new vendor, an open-ended custom project is a harder internal approval than a small, fixed pilot with a clear deliverable.

    High impactLow effort First result in 1 month Co founder & CEO

    First action

    Price the last two engagements per task and set a pilot range.

    How to measure it

    Paid pilots

    Now
    –
    Target
    3
    By
    1 quarter

    Steps

    1. 1

      Cost past work

      Hours and cost per delivered task.

      Strategic Projects Lead 2 days

    2. 2

      Define the pilot

      Scope, deliverables, timeline, credit.

      Co-founder & CEO 2 days

    3. 3

      Write terms

      A short order form alongside the terms page.

      Co-founder & CEO 1 week

    4. 4

      Publish /pricing

      Pilot details and a "talk to us" path for larger work.

      Founding Engineer 1 day

    5. 5

      Offer in outreach

      Lead with the pilot in every first call.

      Co-founder & CEO ongoing

    6. 6

      Track conversion

      Pilot to expansion rate.

      Strategic Projects Lead quarterly

    First result
    1 month
    Cost
    $0
    Owner
    Co founder & CEO

    Smallest test

    Offer the pilot in the next three conversations before publishing it.

    Why it works

    Define one pilot: for example, a set number of long-horizon tasks in one domain with verifiers and a quality report (idea 8), delivered in a fixed number of weeks at a fixed price or price range, credited against a larger contract. Publish it at /pricing with what's included, what the customer provides, and the timeline. Keep larger programmes as "talk to us".

    A priced first step shortens procurement, gives the scoping form a clear offer, and makes it easier for a champion inside a lab to say yes.

    Risks and how to handle them

    • Under-pricing

      Publish a range and credit the pilot against expansion.

    Builds on

    Evidence

    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • HiringStrategic Projects Lead (San Francisco, on-site, posted August 14, 2026): leads cross-functional projects across engineering, operations and go-to-market, owns delivery of RL environments for strategic customers and reports directly to the founders; needs a technical background (at least SQL or Python data analysis), with human data or RL experience preferred.Source
  • 10

    MangoDesk's whole site is one URL with no sitemap. Its own descriptions cover software engineering, real-world workflows and the "$5T knowledge work economy", while RL List organises buyers by category (coding, computer use, enterprise). A buyer searching for environments in a specific workflow has no MangoDesk page to land on.

    Medium impactMedium effort First result in 2-3 months Co founder & CEO

    First action

    List the environment types delivered so far.

    How to measure it

    Monthly visits to environment pages

    Now
    –
    Target
    500
    By
    2 quarters

    Steps

    1. 1

      List types

      Environment types delivered or ready to deliver.

      Strategic Projects Lead 1 day

    2. 2

      Template

      Tasks, horizon, verifiers, sample, form.

      Co-founder & CEO 1 day

    3. 3

      Write three pages

      Starting with the strongest types.

      Co-founder & CEO 2 weeks

    4. 4

      Build

      Routes, sitemap.xml, internal links.

      Founding Engineer 2 days

    5. 5

      Submit

      Search Console and directory profiles.

      Founding Engineer 1 hour

    6. 6

      Expand

      Add types from scoping-form demand.

      Strategic Projects Lead quarterly

    First result
    2-3 months
    Cost
    Writing time
    Owner
    Co founder & CEO

    Smallest test

    Publish the software-engineering page first.

    Why it works

    Publish a page for each environment type MangoDesk actually delivers (for example, software engineering, tool-using agents, and one knowledge-work domain). Each page covers the task types, horizon length, verifier approach, a sample task, and a link to the scoping form. Add them to a new sitemap.xml. Use the domains people request most in the scoping form (idea 6) to decide what comes next.

    Category pages match how buyers and directories search, and each one is a citable primary source.

    Risks and how to handle them

    • Thin pages

      Every page needs a real sample task and verifier.

    Builds on

    Evidence

    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • Y CombinatorYC company page: Summer 2025 batch, team size 2, San Francisco; one-liner "Long-horizon RL environments"; "We help AI become better at real-world workflows through production-grade RL environments."; launch post titled "MangoDesk: Cursor for Evals and Human Data"; links a LinkedIn page and no X account.Source
    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
  • 11

    MangoDesk already built a package that runs tool-use evaluations (retail, airline, telecom) and uploads trajectories to a MangoDesk cloud dashboard. It's a ready-made product-led entry point for agent teams, but it hasn't been released since January 2026 and its docs are unreachable.

    Medium impactMedium effort First result in 1-2 months Founding Engineer

    First action

    Check whether the dashboard still accepts uploads from the current version.

    How to measure it

    Teams uploading runs monthly

    Now
    –
    Target
    15
    By
    2 quarters

    Steps

    1. 1

      Confirm it works

      End-to-end test with current tau-bench.

      Founding Engineer 2 days

    2. 2

      Update

      Release with fixed links and docs.

      Founding Engineer 1 week

    3. 3

      Add the next step

      A dashboard prompt to try a MangoDesk environment.

      Founding Engineer 2 weeks

    4. 4

      Write a guide

      "Evaluate your tool-using agent in 10 minutes".

      Co-founder & CTO 3 days

    5. 5

      Share

      Agent-builder communities and the YC network.

      Co-founder & CEO 2 weeks

    6. 6

      Follow up

      Contact teams with repeated runs.

      Co-founder & CEO weekly

    First result
    1-2 months
    Cost
    Engineering time and hosting
    Owner
    Founding Engineer

    Smallest test

    Release a working version and share it in one community.

    Why it works

    After fixing the links (idea 5), keep the runner free and position it as "run a public benchmark in minutes and see trajectories in the dashboard". In the dashboard, offer a next step: run the same agent on one of MangoDesk's own long-horizon environments (idea 7) or request a custom one. Track which teams install it and reach out to active ones.

    Engineers who already see their agent's trajectories in MangoDesk's dashboard are warm leads for custom environments.

    Risks and how to handle them

    • Users stay on the free tier

      Keep it scoped to public benchmarks; custom environments are the paid offer.

    Builds on

    Evidence

    • PyPIPackage "mangodesk" 0.5.4: "tau-bench wrapper with cloud dashboard and trajectory tracking" that runs evaluations in retail, airline and telecom environments and uploads results to MangoDesk's cloud; proprietary licence; nine releases between December 30, 2025 and January 2, 2026, none since. Its Documentation link (docs.mangodesk.com) does not resolve and its GitHub Repository link returns 404 (checked November 10, 2026).Source
  • 12

    MangoDesk delivers RL environments to strategic customers, and the new Strategic Projects Lead owns that delivery. Environments lose value as models saturate them or learn to exploit them. The press feature flagged retention as a key thing to watch. A one-off delivery ends the relationship, while a refresh programme keeps it going.

    High impactMedium effort First result in 1 quarter Strategic Projects Lead

    First action

    Add a 30-day check-in to every current delivery.

    How to measure it

    Share of customers renewing within 90 days

    Now
    –
    Target
    50%
    By
    2 quarters

    Steps

    1. 1

      Define refresh tiers

      Monthly or quarterly variants and verifier patches.

      Strategic Projects Lead 1 week

    2. 2

      Build tooling

      Generate harder variants from existing tasks.

      Founding Engineer 3 weeks

    3. 3

      Saturation report

      Where the customer's model now scores.

      Founding Engineer 1 week

    4. 4

      Schedule check-ins

      30 days after each delivery.

      Strategic Projects Lead ongoing

    5. 5

      Price it

      Discounted against new builds.

      Co-founder & CEO 2 days

    6. 6

      Track

      Renewal rate and revenue per customer.

      Strategic Projects Lead quarterly

    First result
    1 quarter
    Cost
    Engineering time
    Owner
    Strategic Projects Lead

    Smallest test

    Offer a refresh to the most recent customer.

    Why it works

    Offer every delivered environment with a refresh option: harder task variants every month or quarter, patched verifiers when exploits are found, and a short saturation report showing where the customer's model now scores. Set up a check-in 30 days after each delivery to propose the refresh.

    Recurring revenue from existing lab relationships is cheaper than winning new ones, and the saturation report shows when the customer needs a refresh.

    Risks and how to handle them

    • Refresh work crowds out new builds

      Automate variant generation first.

    Builds on

    Evidence

    • HiringStrategic Projects Lead (San Francisco, on-site, posted August 14, 2026): leads cross-functional projects across engineering, operations and go-to-market, owns delivery of RL environments for strategic customers and reports directly to the founders; needs a technical background (at least SQL or Python data analysis), with human data or RL experience preferred.Source
    • PressHuge Magazine feature (February 2026): names no MangoDesk customers or design partners; says generation speed does not prove output quality and "the actual quality of the generated data is the only metric that matters"; plans to watch generated-pipeline quality against hand-built alternatives, domain edge cases and retention.Source
  • 13

    RL List tracks the execution infrastructure that agent rollouts run on (Modal, Daytona, E2B, Runloop, Morph and others) separately from environment vendors. Labs that already use one of these sandboxes would adopt environments faster if they ran there without extra setup, and these providers want showcase workloads.

    Medium impactMedium effort First result in 2-3 months Co founder & CTO

    First action

    Ask the last three customers which sandbox they run rollouts on.

    How to measure it

    Scoping requests citing a partner

    Now
    –
    Target
    3
    By
    2 quarters

    Steps

    1. 1

      Ask customers

      Which sandbox they use.

      Strategic Projects Lead 1 week

    2. 2

      Pick one provider

      The one most customers use.

      Co-founder & CTO 1 day

    3. 3

      Package

      One environment runnable on it.

      Founding Engineer 2 weeks

    4. 4

      Joint tutorial

      Co-written with the provider's devrel team.

      Co-founder & CTO 2 weeks

    5. 5

      Get listed

      Examples or partner pages.

      Co-founder & CEO 2 weeks

    6. 6

      Measure

      Tagged inbound.

      Co-founder & CEO monthly

    First result
    2-3 months
    Cost
    Engineering time
    Owner
    Co founder & CTO

    Smallest test

    Pitch a joint tutorial to one provider.

    Why it works

    Package one MangoDesk environment so it runs out of the box on one sandbox provider, write a joint tutorial, and ask for a listing in the provider's examples or partner pages. Measure inbound from that channel before adding a second provider.

    Infrastructure partners reach the same engineers MangoDesk sells to, and a ready-to-run environment removes integration friction from a pilot.

    Risks and how to handle them

    • Lock-in to one provider

      Keep environments portable and add a second provider later.

    Evidence

    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
  • 14

    The founders have the background buyers trust: one built a GenAI data engine at Scale AI serving frontier labs, the other previously founded a software firm with seven-figure revenue. Yet the site links no social profiles, YC lists no X account, and the only outreach asset is a one-sentence homepage.

    High impactMedium effort First result in 1 month Co founder & CEO

    First action

    List 30 target teams and the public work each has published.

    How to measure it

    Scoping calls from outbound

    Now
    –
    Target
    6
    By
    1 quarter

    Steps

    1. 1

      Build the list

      30 teams with public post-training or agent work.

      Strategic Projects Lead 1 week

    2. 2

      Prepare samples

      One relevant sample task per domain.

      Strategic Projects Lead 1 week

    3. 3

      Write notes

      Short, specific, pilot offer.

      Co-founder & CEO 2 days

    4. 4

      Send

      Five a day.

      Co-founder & CEO 6 weeks

    5. 5

      Post weekly

      A technical note from a founder account.

      Co-founder & CTO weekly

    6. 6

      Link profiles

      Founder and company profiles in the site footer.

      Founding Engineer 30 minutes

    7. 7

      Review

      Reply and call rates by domain.

      Co-founder & CEO biweekly

    First result
    1 month
    Cost
    Founder time
    Owner
    Co founder & CEO

    Smallest test

    Send ten notes with a sample task and count replies.

    Build the prospect list. The list compiled from public sources, verified, deduplicated and delivered to your CRM. Usually live within days.

    Book a working session

    Why it works

    Build a list of post-training and agent-evaluation leads at labs and agent companies, using public talks and papers. Send short notes from the CEO that include one sample task from the buyer's domain (idea 1) and offer the fixed pilot (idea 9). Alongside, post one technical note a week from the founders' accounts (verifier design, reward hacking found, benchmark results) and link the profiles from the site.

    In a small, technical market, a founder who sends a working sample gets more replies than a sales email.

    Risks and how to handle them

    • Founder time competes with delivery

      Cap outbound at one hour a day.

    Builds on

    Evidence

    • Industry directoryStartup profile (standout.work): 2 employees, Seed stage, "$130K total raised" with Y Combinator as the listed investor; the CEO was a software engineer at Scale AI who helped build a GenAI data engine serving frontier labs; the CTO previously founded a software firm with seven-figure revenue serving Fortune 500 clients.Source
    • Your websiteSite build (programmable.info scan, November 10, 2026): the homepage is built with Lovable (meta author "Lovable", images under /lovable-uploads/) and hosted on Render behind Cloudflare; no analytics, blog, newsletter, live chat or social profile links detected.Source
    • Y CombinatorYC company page: Summer 2025 batch, team size 2, San Francisco; one-liner "Long-horizon RL environments"; "We help AI become better at real-world workflows through production-grade RL environments."; launch post titled "MangoDesk: Cursor for Evals and Human Data"; links a LinkedIn page and no X account.Source
  • 15

    AI labs send vendors security questionnaires before sharing models or data. MangoDesk's /trust page is a 404, RL List scores vendors on security posture such as SOC 2, and rivals such as Proximal publish an MSA alongside their privacy policy. A two-person vendor without a security pack can lose a deal it has already won technically.

    High impactMedium effort First result in 1 quarter Co founder & CTO

    First action

    Fill in a standard security questionnaire with current answers.

    How to measure it

    Procurement cycle time

    Now
    –
    Target
    -30%
    By
    2 quarters

    Steps

    1. 1

      Questionnaire

      Answer a standard one honestly.

      Co-founder & CTO 2 days

    2. 2

      Close gaps

      Access, logging, backups, isolation.

      Founding Engineer 3 weeks

    3. 3

      Templates

      MSA, DPA, subprocessors list.

      Co-founder & CEO 2 weeks

    4. 4

      Start SOC 2

      Readiness with a compliance tool.

      Co-founder & CTO 1 quarter

    5. 5

      Publish /trust

      Current controls and status.

      Founding Engineer 1 day

    6. 6

      Use in deals

      Send the pack with every scope.

      Strategic Projects Lead ongoing

    First result
    1 quarter
    Cost
    Compliance tooling and audit fees
    Owner
    Co founder & CTO

    Smallest test

    Send the filled questionnaire to the next prospect before they ask.

    Why it works

    Assemble a security pack: a data-handling and isolation overview, an access-control policy, subprocessors, an incident contact, an MSA and DPA template, and filled-in answers to a standard questionnaire. Start SOC 2 Type I readiness with a compliance-automation tool and say so honestly on /trust ("Type I in progress, expected <quarter>").

    Security review is the longest step in selling to labs. A ready pack shortens it, and the trust page becomes a primary source for directories.

    Risks and how to handle them

    • Cost for a seed-stage team

      Start with the pack and Type I; pursue Type II when a deal needs it.

    Builds on

    Evidence

    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
    • Competitor researchRivals' own homepages: Proximal leads with its FrontierSWE v2 benchmark ("34 tasks, 13 frontier models, 20-hour budgets"), a public leaderboard, a blog and Privacy and MSA pages; Halluminate leads with an 88-problem finance due-diligence benchmark "written and reviewed by practicing finance deal professionals" and has Research, Expert Network and Data Partnerships pages; Datacurve promotes its "DeepSWE" long-horizon coding benchmark and a research page; Fleet links Privacy, Terms, Cookies and a status page.Source
  • 16

    Long-horizon environments for knowledge work need practitioners to write and review tasks. Halluminate's finance benchmark is "written and reviewed by practicing finance deal professionals", and its site has a dedicated Expert Network page. MangoDesk's Strategic Projects Lead post asks for human-data experience, but the site's only recruiting message is for full-time engineers, researchers and operators.

    Medium impactMedium effort First result in 1 month Strategic Projects Lead

    First action

    Write the expert page for the first target domain.

    How to measure it

    Active vetted experts

    Now
    –
    Target
    50
    By
    1 quarter

    Steps

    1. 1

      Define roles

      Author and reviewer, by domain.

      Strategic Projects Lead 2 days

    2. 2

      Set pay and vetting

      Rates, test task, NDA.

      Strategic Projects Lead 3 days

    3. 3

      Publish /experts

      With an application form.

      Founding Engineer 2 days

    4. 4

      Referral links

      Bonus after the referred expert's first batch.

      Founding Engineer 1 week

    5. 5

      Seed

      Invite current contributors to refer.

      Strategic Projects Lead 1 week

    6. 6

      Track quality

      Review pass rate by source.

      Strategic Projects Lead monthly

    First result
    1 month
    Cost
    Referral bonuses
    Owner
    Strategic Projects Lead

    Smallest test

    Ask current contributors for two referrals each.

    Why it works

    Publish a /experts page for paid, part-time task authors and reviewers in the domains MangoDesk targets. Explain the work, pay ranges, vetting and confidentiality. Give each accepted expert a referral link with a bonus once a referred expert completes their first paid batch. Track quality by referral source.

    Experts know other experts. Referral supply is cheaper and better vetted than job-board recruiting, and a deep bench is something buyers ask about.

    Risks and how to handle them

    • Low-quality referrals

      Pay the bonus only after a vetted first batch.

    Evidence

    • Competitor researchRivals' own homepages: Proximal leads with its FrontierSWE v2 benchmark ("34 tasks, 13 frontier models, 20-hour budgets"), a public leaderboard, a blog and Privacy and MSA pages; Halluminate leads with an 88-problem finance due-diligence benchmark "written and reviewed by practicing finance deal professionals" and has Research, Expert Network and Data Partnerships pages; Datacurve promotes its "DeepSWE" long-horizon coding benchmark and a research page; Fleet links Privacy, Terms, Cookies and a status page.Source
    • HiringStrategic Projects Lead (San Francisco, on-site, posted August 14, 2026): leads cross-functional projects across engineering, operations and go-to-market, owns delivery of RL environments for strategic customers and reports directly to the founders; needs a technical background (at least SQL or Python data analysis), with human data or RL experience preferred.Source
    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
  • 17

    /changelog returns 404 and the site shows no dates other than the copyright, so a visitor can't tell whether MangoDesk is shipping. Its public package hasn't been released since January 2026. For a seed-stage vendor, visible momentum is part of the sale.

    Low impactLow effort First result in 2 weeks Co founder & CTO

    First action

    Write the entry covering the last three months.

    How to measure it

    Changelog entries

    Now
    –
    Target
    1 a month
    By
    ongoing

    Steps

    1. 1

      Back-fill

      One entry covering the last quarter.

      Co-founder & CTO 2 hours

    2. 2

      Build the page

      /changelog with dates.

      Founding Engineer 2 hours

    3. 3

      Footer link

      From every page.

      Founding Engineer 15 minutes

    4. 4

      Monthly entry

      Environments, verifiers, tools, team.

      Co-founder & CTO monthly

    5. 5

      Reuse

      As a founder post.

      Co-founder & CEO monthly

    6. 6

      Check

      No customer names or confidential detail.

      Strategic Projects Lead monthly

    First result
    2 weeks
    Cost
    $0
    Owner
    Co founder & CTO

    Smallest test

    Publish the back-filled entry.

    Why it works

    Publish a short monthly entry at /changelog covering new environment types, verifier improvements, benchmark updates, tooling releases and team hires. Never name customers. Link it from the footer and reuse each entry as a founder post (idea 14).

    It's a cheap signal that the company is active, and it gives directories and buyers dated primary sources.

    Risks and how to handle them

    • Gaps look worse than no changelog

      Commit to monthly, even if short.

    Evidence

    • Your websitePages checked November 10, 2026 (privacy checked twice): /privacy, /terms, /trust, /pricing and /changelog all show the site's "404 Oops! Page not found" page, and /sitemap.xml returns 404. robots.txt allows every crawler.Source
    • PyPIPackage "mangodesk" 0.5.4: "tau-bench wrapper with cloud dashboard and trajectory tracking" that runs evaluations in retail, airline and telecom environments and uploads results to MangoDesk's cloud; proprietary licence; nine releases between December 30, 2025 and January 2, 2026, none since. Its Documentation link (docs.mangodesk.com) does not resolve and its GitHub Repository link returns 404 (checked November 10, 2026).Source

Big bets 3

Larger plays that can change the trajectory.

  • 18

    MangoDesk launched as "Cursor for Evals and Human Data", spinning up "bespoke data-annotation pipelines in seconds", and the meta description still says "evals and data pipelines with natural language". Today the company sells custom RL environments through founder conversations, and there's no way to try the product. A self-serve tier would reach the many agent teams that will never sign a custom contract.

    High impactHigh effort First result in 1 quarter Co founder & CTO

    First action

    Add a waitlist for self-serve evals and count sign-ups for a month.

    How to measure it

    Paying self-serve teams

    Now
    –
    Target
    25
    By
    3 quarters

    Steps

    1. 1

      Waitlist

      One section and a form on the site.

      Founding Engineer 1 day

    2. 2

      Interview

      Ten waitlist teams about their eval workflow.

      Co-founder & CEO 3 weeks

    3. 3

      Scope the MVP

      Describe, generate, run, export.

      Co-founder & CTO 1 week

    4. 4

      Build

      On the existing pipeline-generation code.

      Founding Engineer 2 months

    5. 5

      Beta

      Ten teams, free for a month.

      Co-founder & CTO 1 month

    6. 6

      Price

      Per seat or per item.

      Co-founder & CEO 2 weeks

    7. 7

      Route upsells

      RL-heavy requests to the environments team.

      Strategic Projects Lead ongoing

    First result
    1 quarter
    Cost
    Two quarters of engineering
    Owner
    Co founder & CTO

    Smallest test

    Waitlist plus ten interviews.

    Why it works

    Ship a gated self-serve workspace: a team describes the eval in plain language, MangoDesk generates the task schema, rubric and annotation pipeline, and the team runs it with its own reviewers or requests MangoDesk experts. Price it per seat or per evaluated item, and route large or RL-heavy requests to the environments team. Start with a waitlist on the site to measure demand before building.

    It gives MangoDesk a second, scalable revenue line, and it feeds the custom-environment business with qualified leads.

    Risks and how to handle them

    • Splits a two-person team's focus

      Gate on waitlist demand and hire before building.

    Builds on

    Evidence

    • PressLaunch listing (August 13, 2025): "MangoDesk: Cursor for Evals and Human Data", to "Accelerate creating custom evals and post-training data by spinning up bespoke data-annotation pipelines in seconds"; company line "Simple platform to create data for AI."Source
    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • PressHuge Magazine feature (February 2026): names no MangoDesk customers or design partners; says generation speed does not prove output quality and "the actual quality of the generated data is the only metric that matters"; plans to watch generated-pipeline quality against hand-built alternatives, domain edge cases and retention.Source
  • 19

    MangoDesk's mission targets the "$5T knowledge work economy", but most of the RL-environment field clusters around coding and computer use. Halluminate shows that owning one vertical works: it leads with finance and consulting and an expert-written benchmark. A two-person team can't cover all of knowledge work, but it can become the default vendor for one workflow family.

    High impactHigh effort First result in 1-2 quarters Co founder & CEO

    First action

    Map which knowledge-work verticals RL List vendors already cover.

    How to measure it

    Share of new revenue in the vertical

    Now
    –
    Target
    60%
    By
    3 quarters

    Steps

    1. 1

      Map the field

      Verticals covered by listed vendors.

      Strategic Projects Lead 1 week

    2. 2

      Match demand

      Scoping-form and outbound requests.

      Strategic Projects Lead 1 week

    3. 3

      Choose

      One vertical with demand and few rivals.

      Co-founder & CEO 1 day

    4. 4

      Build the suite

      Flagship environments and verifiers.

      Co-founder & CTO 1 quarter

    5. 5

      Recruit experts

      Practitioners for task authoring.

      Strategic Projects Lead 1 month

    6. 6

      Reposition

      Site, profiles and outreach.

      Co-founder & CEO 2 weeks

    7. 7

      Review

      Win rate in the vertical against others.

      Co-founder & CEO quarterly

    First result
    1-2 quarters
    Cost
    Expert fees and two quarters of team focus
    Owner
    Co founder & CEO

    Smallest test

    Pitch a vertical pilot to three prospects before building.

    Why it works

    Use scoping-form demand (idea 6) and the founders' networks to pick one vertical, such as operations, sales operations or analytics work, that is underserved by the vendors on RL List. Build a flagship environment suite, a public benchmark (idea 7) and an expert bench (idea 16) for it, and position the whole site around it for two quarters before adding a second vertical.

    Focus builds depth that generalists can't match, and labs buy from the vendor with the best environments in the capability they're trying to improve.

    Risks and how to handle them

    • Picking a vertical labs don't prioritise

      Choose from real requests, not guesses.

    Builds on

    Evidence

    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • Competitor researchRivals' own homepages: Proximal leads with its FrontierSWE v2 benchmark ("34 tasks, 13 frontier models, 20-hour budgets"), a public leaderboard, a blog and Privacy and MSA pages; Halluminate leads with an 88-problem finance due-diligence benchmark "written and reviewed by practicing finance deal professionals" and has Research, Expert Network and Data Partnerships pages; Datacurve promotes its "DeepSWE" long-horizon coding benchmark and a research page; Fleet links Privacy, Terms, Cookies and a status page.Source
    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
    • HiringFounding Engineer (San Francisco, on-site, posted February 9, 2026): MangoDesk is a YC-backed startup building data infrastructure and environments to connect AI with knowledge work; the role needs production backend or full-stack systems, scalability and end-to-end ownership, with applied AI or RL post-training experience preferred.Source
  • 20

    RL List tracks HUD, "an SDK and marketplace for defining and selling RL environments", and open environment hubs such as Prime Intellect's. Buyers already browse these to find environments, and listing there is a sales channel that doesn't depend on MangoDesk's own traffic.

    Medium impactMedium effort First result in 1-2 months Co founder & CTO

    First action

    Read the listing terms of one marketplace and one open hub.

    How to measure it

    Custom requests from listings

    Now
    –
    Target
    5
    By
    2 quarters

    Steps

    1. 1

      Review terms

      Licensing, revenue share, attribution.

      Co-founder & CEO 1 week

    2. 2

      Pick environments

      Non-proprietary, representative ones.

      Co-founder & CTO 2 days

    3. 3

      Package

      To the hub's specification.

      Founding Engineer 2 weeks

    4. 4

      List

      With a link to request custom versions.

      Founding Engineer 1 day

    5. 5

      Announce

      Founder posts and the changelog.

      Co-founder & CEO 1 day

    6. 6

      Track

      Downloads and requests.

      Strategic Projects Lead monthly

    First result
    1-2 months
    Cost
    Engineering time; possible revenue share
    Owner
    Co founder & CTO

    Smallest test

    List one environment on one hub.

    Why it works

    Publish one or two non-proprietary MangoDesk environments, or a free sample tier, on an environment marketplace or hub, with clear attribution and a link to request custom or harder versions. Track downloads and which teams ask for more. Keep flagship and customer-specific environments private.

    Marketplaces put MangoDesk in front of teams already shopping for environments, and a free sample shows quality before procurement.

    Risks and how to handle them

    • Giving away value

      Share only samples; keep harder variants and verifiers paid.

    Evidence

    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source

Wildcards 2

Unexpected experiments, each with a cheap way to test it.

  • 21

    The homepage says MangoDesk "recently raised an oversubscribed seed round led by top-tier VCs", but a public startup profile still shows "$130K total raised" with Y Combinator as the only investor. RL List's score weights funding raised. If the round isn't public anywhere, every directory and buyer due-diligence check will understate the company.

    Medium impactLow effort First result in 1 month Co founder & CEO

    First action

    Agree with the lead investor what can be disclosed.

    How to measure it

    Directories with the updated round

    Now
    –
    Target
    5
    By
    60 days

    Steps

    1. 1

      Agree disclosure

      Amount or range, leads.

      Co-founder & CEO 1 week

    2. 2

      Write the post

      What the round funds.

      Co-founder & CEO 1 day

    3. 3

      Publish

      On mangodesk.com with a date.

      Founding Engineer 1 hour

    4. 4

      Update profiles

      YC, LinkedIn, startup databases.

      Co-founder & CEO 1 day

    5. 5

      Pair with news

      Benchmark or directory launch.

      Co-founder & CEO 2 weeks

    6. 6

      Check

      Directories reflect the update.

      Strategic Projects Lead monthly

    First result
    1 month
    Cost
    $0
    Owner
    Co founder & CEO

    Smallest test

    Update the YC and LinkedIn profiles with what can be said today.

    Why it works

    If the investors agree, publish a short announcement on mangodesk.com (amount or range, lead investors, what the money builds) and update the YC profile, LinkedIn and startup databases. Use the announcement as the hook for the benchmark launch (idea 7) and the directory profile (idea 4). If the round can't be disclosed, soften the homepage line so it matches what is public.

    A visible raise reassures lab procurement that the vendor will be around for a multi-quarter programme, and directories then rank it more accurately.

    Risks and how to handle them

    • Investors prefer silence

      Align the homepage wording with what is public instead.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): one page whose only calls to action are "Careers", "Contact us" and "email [email protected]". H1 "Measurable model improvement": "We evaluate and improve AI on meaningful use cases via production grade RL environments." The title says "Accelerate custom evals and data" and the meta description "MangoDesk helps you create production-grade evals and data pipelines with natural language." Proof shown: a Y Combinator "Backed by" logo, a team "with experience from Scale AI, Uber", "an oversubscribed seed round" and the mission "to bridge the gap between AI and the $5T knowledge work economy through data". No customers, case studies, environment examples or pricing.Source
    • Industry directoryStartup profile (standout.work): 2 employees, Seed stage, "$130K total raised" with Y Combinator as the listed investor; the CEO was a software engineer at Scale AI who helped build a GenAI data engine serving frontier labs; the CTO previously founded a software firm with seven-figure revenue serving Fortune 500 clients.Source
    • Industry directoryRL List (snapshot dated October 5, 2026) ranks and tracks 42 companies that build RL environments for frontier labs, among them Proximal, Halluminate, Datacurve, Fleet AI, HUD (an SDK and marketplace for selling RL environments) and sandbox providers such as Modal, Daytona and E2B. MangoDesk is not listed. Its methodology scores funding, customer traction, named or verified customers, frontier-lab ties and security posture such as SOC 2, plus a bonus for each field backed by a primary source.Source
  • 22

    The scan found no analytics on mangodesk.com; the only third-party request is one small AWS asset. Without analytics MangoDesk can't tell how many lab visitors arrive, which pages they read, or whether the buyer page, scoping form or benchmark change anything.

    Medium impactLow effort First result in 1 week Founding Engineer

    First action

    Install the script and define the scoping-form goal.

    How to measure it

    Conversion goals live

    Now
    –
    Target
    3
    By
    2 weeks

    Steps

    1. 1

      Choose a tool

      Cookieless and lightweight.

      Founding Engineer 1 hour

    2. 2

      Install

      One script in the Lovable site.

      Founding Engineer 30 minutes

    3. 3

      Define goals

      Form submit, pilot view, docs visit.

      Founding Engineer 1 hour

    4. 4

      Tag links

      UTMs on outbound, PyPI and directory links.

      Founding Engineer 1 hour

    5. 5

      Update privacy page

      Mention the analytics tool.

      Co-founder & CTO 15 minutes

    6. 6

      Review

      Monthly against plan metrics.

      Co-founder & CEO monthly

    First result
    1 week
    Cost
    $0-20 a month
    Owner
    Founding Engineer

    Smallest test

    Install today and check visits after a week.

    Why it works

    Add a privacy-friendly, cookieless analytics tool (so no consent banner is needed in most cases), set three goals (scoping-form submit, pilot page view, docs visit), add UTM tags to outbound, package and directory links, and review the numbers monthly against this plan's metrics.

    It costs almost nothing and turns every other idea from guesswork into a measured test.

    Risks and how to handle them

    • Low traffic makes numbers noisy

      Track absolute counts of scoping requests, not rates.

    Evidence

    • Your websiteSite build (programmable.info scan, November 10, 2026): the homepage is built with Lovable (meta author "Lovable", images under /lovable-uploads/) and hosted on Render behind Cloudflare; no analytics, blog, newsletter, live chat or social profile links detected.Source
    • PageSpeedGoogle PageSpeed, November 10, 2026: mobile performance 72 (desktop 84); Largest Contentful Paint 30.5 s on mobile and 2.6 s on desktop, the LCP element being the "Measurable model improvement" text block; the page weighs 5,957 KB, 5,738 KB of it images, across 10 requests; "Improve image delivery" could save 5,222 KB and longer cache lifetimes 2,821 KB. Accessibility, best practices and SEO all score 100. There is no real-visitor data yet, so these are lab results.Source

03

How Programmable can help

Where an idea in this playbook is something Programmable builds or runs, it's listed here with what we'd do. Everything else is yours to run.

Have it built

ProgrammableAgency

1 of the 22 ideas are things Programmable builds and runs for its customers: pages, outreach, prospect lists, booking flows and monitors. Pick one, or send the whole list.

Prospect lists1The list compiled from public sources, verified, deduplicated and delivered to your CRM.

A 30-minute session with the engineer who would build it, not a sales call. You leave with a plan whether or not you hire us.

Book a working session

Usually live within days.

Tools for the rest

ProgrammableInfo

Turn the research into a list you can act on

These ideas start with finding the right companies. Info finds them and the right person at each, researches every one and drafts the outreach, then asks before it sends.

  • Build this list in Info for idea #14 (opens programmable.info in a new tab)

New accounts: 100,000 free credits with code BETA100K

Find the companies (opens programmable.info in a new tab)
More example playbooks and reviews

More growth playbooks

Other companies we have researched, each read from what they are hiring for.

  • Go Nimbly growth strategyA revenue operations (RevOps) consultancy that designs and builds go-to-market systems: CRM architecture, quote-to-cash, lead routing…23 growth ideas
  • Groupon growth strategyA local-commerce marketplace for discounted experiences, services, travel and goods, moving from daily deals to an experience discovery…20 growth ideas
  • SandboxAQ growth strategyBuilds Large Quantitative Models (physics- and chemistry-grounded AI) for drug discovery and materials, plus cybersecurity (AQtive Guard)…22 growth ideas
  • Workato growth strategyA control and execution platform for enterprise AI, built on its iPaaS: integrations and automation, agents (AIRO, Agent Studio)…21 growth ideas
  • CareSwift growth strategyAI-assisted EMS narrative software for agencies. Crews enter call details, answer clarifying questions built from the agency's own…23 growth ideas
  • Recorded Future growth strategyA threat intelligence company: the Intelligence Graph and Insikt Group research power four solutions (Cyber Operations, Digital Risk…22 growth ideas
  • Solaris growth strategyA Berlin-based, fully licensed German bank offering Banking-as-a-Service: accounts, cards, payments, lending and KYC via API, including…21 growth ideas
  • Emergent growth strategyAn AI app builder ("vibe coding") where autonomous agents design, code, test and deploy web and mobile apps from plain-language prompts…22 growth ideas

See every playbook and website review

Independent research. Programmable wrote this playbook from public sources, as they stood on November 10, 2026, and cites them throughout. MangoDesk did not commission, review or endorse it, and Programmable is not affiliated with MangoDesk. Company and product names are trademarks of their owners, used only to identify them. The ideas are suggestions and opinion, not professional, legal or financial advice, and not statements about how MangoDesk runs its business.

Something wrong, or do you represent MangoDesk? Email [email protected] and we will correct it or take it down.

30/60/90 roadmap

What to focus on in each month, and the ideas that carry it.

  1. Days 1 to 307 ideas

    Make the site buyable: buyer page, legal and trust pages, image fix, one positioning line, working package links, a scoping form, analytics and the RL List profile.

  2. Days 31 to 6010 ideas

    Build proof and a sales motion: public benchmark, quality reports, fixed-scope pilot, environment pages, the open eval runner, refresh contracts, founder outbound, the trust pack, the expert page and the changelog.

  3. Days 61 to 905 ideas

    Commit to bigger bets: sandbox partners, self-serve pipelines, a vertical focus, marketplace distribution and an accurate public funding record.

The research behind the plan

3 research areas, each summarised in a few lines. Open any of them for the full detail.

Customers

What customers say

Customers value measurable model improvement on meaningful, real-world use cases. The most cited problem: the eval and post-training data bottleneck.

  • Production-grade, long-horizon environments rather than toy tasks
  • Doubts about whether generated data is high quality
  • In their words: “RL environments”

Market

Competitive landscape

5 competitors compared: Proximal, Halluminate, Datacurve, Fleet AI and 1 more.

  • Gap to exploit: Environments for broad knowledge work beyond coding and finance
  • Proximal: Coding-centred
  • Halluminate: Narrow first vertical

Hiring signals

MangoDesk is a two-person team hiring its first two people, both on-site in San Francisco.

  • Strategic Projects Lead: Owns delivery of RL environments for strategic customers. Ideas 6, 8, 9, 12, 16.
  • Founding Engineer: Data infrastructure and environments for knowledge work. Ideas 2, 5, 11, 22.

From the team that wrote this playbook

All products
  • ProgrammableAgencyWorking AI systems in days, not quarters (opens programmable.agency in a new tab)
  • ProgrammableInfoAn agent that finds, researches and asks first (opens programmable.info in a new tab)
  • ProgrammableDigitalScreenshots in, an animated product demo out (opens programmable.digital in a new tab)
  • ProgrammableWatchKnow the day a supplier falls out of compliance (opens programmable.watch in a new tab)
  • ProgrammableTalkGet registrants for your events, and know who’s coming (opens programmable.talk in a new tab)
  • ProgrammablePageText what you need, your customer gets a page that works (opens programmable.page in a new tab)
  • ProgrammableCommunityForums with labelled AI members, never empty (opens programmable.community in a new tab)
ProgrammableMarketing
ExamplesLeaderboardsProductsTermsPrivacyCookiesDPALegalContact

© 2026 AI-Powered Growth Intelligence