Ambral growth playbook

Publish documents for the product Ambral sells today. Write a Data Use page covering what is stored from each request, how long traces are kept, who at Ambral can see them, that a customer's traffic trains only that customer's own models, and what happens on deletion. Update the privacy policy so it names the endpoint and trace capture. Add service terms with an uptime target and support hours, plus a short subprocessor list. Link all of it from the homepage footer, the book page and the trust center, and send it with every booked call so the champion can forward it to security before the first meeting.

For a buyer, routing production prompts through a new vendor is a data decision before it is a cost decision. When the documents match the product, the main reason a champion can't get approval goes away, and the security review after a good demo gets shorter. Rivals with years of enterprise paperwork set the bar buyers compare against.

[ ] What each request stores (prompt, output, metadata)
[ ] How long traces are kept, and how a customer shortens it
[ ] Who at Ambral can see traces, and when
[ ] A customer's traffic trains only that customer's models
[ ] Closed models: evaluation and inference only
[ ] Deletion on request and at contract end
[ ] Subprocessors and regions
[ ] Uptime target and support hours
[ ] Link to the trust center
[ ] Version and date

• Promising retention or isolation terms the system can't yet enforce

  Publish only what is true today, date the page and update it as features ship.

• Existing enterprise customers signed under the old terms

  Their Enterprise Agreements stay in force; the new terms cover new sign-ups.

• Your websitePrivacy policy (v1.0.4, updated November 2, 2026): describes Ambral as "a cloud-based, AI-powered account management platform", the "Cortex Platform", that helps businesses improve customer relationships through data analysis, automated workflows and predictive insights. It does not describe an inference endpoint or trace capture, gives no fixed retention periods, and says aggregated or de-identified data may be used to train AI models (opt-out by email), while data processed under a customer's Data Processing Addendum is not used for training.Source
• Your websiteTerms of Service (v1.3.0, dated April 20, 2026): the page description reads "AI for account management and revenue growth"; the terms cover "the Site" and services made available through it, with no uptime commitment or SLA ("We have no obligation to provide you with support or maintenance for the Site"); an Enterprise Agreement overrides them.Source
• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source

Rewrite the YC tagline and description around the endpoint ("Inference that learns from your workloads"), update team size and links, and post a new launch on YC that shows the one-line switch, the eval loop and one concrete result. Then ask the main startup directories to refresh their entries, and update the company's social and job-board bios to match.

The YC network is full of AI-native companies, Ambral's stated buyer. A fresh launch reaches them directly, and a consistent description across profiles stops search and AI assistants from repeating the old pitch.

• Existing account-management customers are confused

  Tell them directly first and explain what changes for them.

• Y Combinator profileYC company page: tagline "AI for account management and customer success"; Summer 2025 batch, founded 2025, team size 4, New York. The only launch post is "Ambral Cortex: Enterprise-grade AI Account Managers". Search results and startup directories still describe Ambral as AI account managers for post-sale revenue teams.Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source

Add a short "Your data" section to the homepage and the book page with four answers: what is stored, how long, who trains on it (only you, for your models), and where it runs. Link the trust center and the Data Use page from idea 1, and show compliance status honestly, including anything in progress.

Technical buyers screen vendors before they book. Answering the data question on the page turns a silent drop-off into a booked call, and the call can spend its 20 minutes on the workload.

• Overstating compliance

  Show only certifications held and mark the rest as in progress.

• #1 Rewrite the privacy policy and terms for the inference endpoint: today they describe the Cortex account-management product

• Your websiteTrust center: a Vanta-hosted trust center exists at trust.ambral.com, but the homepage and book page copy say nothing about security, compliance, data retention or where models run.Source
• Your websiteBook page: "Bring a workload. We'll take it from there." A 20-minute walkthrough with an Ambral engineer; visitors bring "one agent or workflow" and its approximate request volume. Booking is a Calendly scheduler (or email); the page itself asks nothing about the workload.Source

Publish a small docs site: a quickstart showing the one-line switch in the common client libraries, how traces appear, how evals are built and reviewed, how routing decisions are made and reversed, rate limits, errors and the models list. Link it from the homepage navigation and every "Chat with us" button.

Engineers trust what they can read. Docs let a champion check the integration and bring a concrete workload to the call, and they are indexed pages that answer the searches developers make.

• Docs drift from the product

  Generate the reference from the API definition.

• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source

Add three questions to the Calendly booking: the workload type (pick from the homepage examples or other), monthly request volume, and current model and monthly spend band. The confirmation email includes a one-page checklist: what a good sample of 50 traces looks like, how to redact it, and what success means for this workload. Route high-volume bookings to the CTO.

Calls that start with the buyer's own numbers end with a concrete next step, and the answers show which workloads and spend levels convert best.

• Extra questions reduce bookings

  Keep it to three, all multiple choice.

• Your websiteBook page: "Bring a workload. We'll take it from there." A 20-minute walkthrough with an Ambral engineer; visitors bring "one agent or workflow" and its approximate request volume. Booking is a Calendly scheduler (or email); the page itself asks nothing about the workload.Source
• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source

Add a short "How pricing works" section to /models (and a /pricing page): what is billed at the listed token rates, what the platform or training fee covers, minimums, and an example net saving for one workload. If prices are negotiated, publish the structure and a starting point, or test a share-of-savings option that ties the fee to the cost reduction Ambral proves.

Cost is the reason buyers come, so pricing is part of the pitch. A clear structure filters out poor-fit leads and gives champions a number for their budget request.

• Public prices anchor enterprise deals too low

  Publish the structure and a starting point, with enterprise terms on request.

• Your websiteModels page: one price table "as of November 5, 2026", in USD per 1M tokens (input, cached input, output), for 16 entries: seven open-weight models available for "evaluation, inference, and training, subject to license" and nine closed-model entries from two providers for "evaluation and inference only, no training", plus cache-write billing notes. There are no per-model pages, comparisons or change history.Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source

Generate one page per model from the same data: prices, cached-input and cache-write rules, whether Ambral can train on it, and the workloads where it tends to match larger models (once idea 12's method is public). Add a price change log, a sitemap that lists every page, and a call to action on each: "Test this model on your own traces".

Price and capability lookups are frequent, high-intent searches by exactly the engineers who manage AI spend, and the data is already maintained. Each page becomes a door into the product.

• Stale prices damage trust

  Update from the same source the billing system uses and show the date.

• Your websiteModels page: one price table "as of November 5, 2026", in USD per 1M tokens (input, cached input, output), for 16 entries: seven open-weight models available for "evaluation, inference, and training, subject to license" and nine closed-model entries from two providers for "evaluation and inference only, no training", plus cache-write billing notes. There are no per-model pages, comparisons or change history.Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source

Keep the training opt-out, but review the list user agent by user agent: allow the crawlers assistants use to fetch pages for live answers and search, and keep blocking bulk training crawlers. Add a short llms.txt that points to the homepage, docs, models and Data Use pages. Re-check the list each quarter, since Cloudflare's managed list changes.

Being cited when a developer asks how to lower AI spend is cheap acquisition for a company with no blog yet. A deliberate setting gets that visibility without giving up the training opt-out.

• A retrieval agent's content is also used for training

  Follow each operator's published documentation and keep the content signal set to no training.

• Your websiterobots.txt is Cloudflare's managed content-signals file: search indexing is allowed, "ai-train=no" is set, and about 30 AI crawlers are fully disallowed.Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source

Add a "How we prove quality" block under the Evaluate step: graders built from the customer's own outcomes, replay of real traces against candidate models, held-out evals that resist overfitting, and the rule that traffic moves only when evals match. Use the same words in the page description and the YC profile.

Every rival promises lower cost; few explain how quality is protected. The team's own description of its engineering is a sharper differentiator than the current headline copy.

• Too technical for budget owners

  Lead with the rule (traffic moves only when quality matches) and put detail below.

• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source
• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source

Agree with one customer to publish an anonymised result: workload type, request volume, starting model and cost, eval method, the share of traffic moved, the cost after, and quality before and after. Put it in place of the illustrative chart, with a one-page write-up and the customer's industry. Aim for three such results by quarter end.

Buyers discount illustrative numbers. One real before-and-after with a stated method is the proof that moves a cost-cutting project up a team's list, and it matches the logo-led proof rivals show.

• Customers won't allow publication

  Offer full anonymity, an industry label only, and review rights.

• Your websiteHomepage numbers, all labelled "Illustrative" or "sample data": cost per 1k requests at matched eval score, an Ambral-trained 27B model at $0.40 (eval 0.95) against a frontier closed model at $6.00 (eval 0.94) and an untuned open 27B model at $0.50 (eval 0.86); a traffic-share chart by model class for data-extraction, claims-intake and code-review. Example workloads shown: research-agent, code-review, claims-intake, contract-review and sales-outreach.Source
• HiringThree open roles, all in New York (secondary San Francisco): Head of Research ($250K–$400K, posted September 1, 2026), Member of Technical Staff, Evals & Post-Training ($165K–$325K, September 1) and Member of Technical Staff, New Grad ($140K–$175K, August 17). The posts say Ambral "helps enterprises own the intelligence behind their most important workflows" and "We graduated from YC S2025, raised millions in funding, and are already deployed within multi-billion dollar enterprises.", giving enterprises open-weight models they own and control rather than rent.Source

Build a calculator on /models: choose the current model, enter monthly input, cached and output tokens (or monthly spend), pick a workload type, and see today's cost, the cost if a share of traffic moved to a smaller or trained model at matched quality (using ranges from real results), and Ambral's fee from idea 6. End with "Check this on your traces" leading to the booking or shadow-mode path.

It turns a passive price list into a personal number, collects qualified leads with their spend band, and gives champions a figure for their budget conversation.

• Estimates overpromise

  Show ranges, label them as estimates, and link the method.

• #6 Say how Ambral charges, even as a range, next to the model price table

• Your websiteModels page: one price table "as of November 5, 2026", in USD per 1M tokens (input, cached input, output), for 16 entries: seven open-weight models available for "evaluation, inference, and training, subject to license" and nine closed-model entries from two providers for "evaluation and inference only, no training", plus cache-write billing notes. There are no per-model pages, comparisons or change history.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source

Publish a technical write-up of how Ambral builds evals from production traces, designs graders, guards against overfitting and decides when traffic moves. Release a small open artifact with it (a grader template or a replay harness for one public task). Give it a permanent /research page, and present it at one engineering meetup in New York.

A checkable method turns a marketing claim into an engineering argument, earns links and discussion among practitioners, and helps recruit the research team the posts describe.

• Giving away the method

  Publish the principles and one example; keep the production system private.

• The Head of Research is not yet hired

  The CTO and evals engineer start it; the new hire takes it over.

• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source
• Your websiteHomepage numbers, all labelled "Illustrative" or "sample data": cost per 1k requests at matched eval score, an Ambral-trained 27B model at $0.40 (eval 0.95) against a frontier closed model at $6.00 (eval 0.94) and an untuned open 27B model at $0.50 (eval 0.86); a traffic-share chart by model class for data-extraction, claims-intake and code-review. Example workloads shown: research-agent, code-review, claims-intake, contract-review and sales-outreach.Source
• Your websiteSite history and structure: since November 2025 the homepage headline changed from "Your AI Account Management Army" to "Inference that learns from your workloads." and the site was rebuilt on Next.js, with PostHog analytics. Checked November 9, 2026: /sitemap.xml, /docs and /pricing return 404, and there is no blog.Source

Split the homepage below the hero into two paths. "For AI products": the quickstart, the calculator and shadow mode, with a call booked only when the team wants one. "For enterprises": owned models, forward-deployed engineers, security answers and a scoped pilot. Give each its own booking type and follow-up, and track them separately.

Startups want to try before they talk; enterprises want a partner and paperwork. Matching the path to the buyer raises conversion on both sides and stops engineers spending enterprise-style effort on small accounts.

• Splitting focus too early

  Let the call data decide which path gets the hero.

• #4 Publish a quickstart for the one-line switch, and link docs from the homepage

• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source
• HiringThree open roles, all in New York (secondary San Francisco): Head of Research ($250K–$400K, posted September 1, 2026), Member of Technical Staff, Evals & Post-Training ($165K–$325K, September 1) and Member of Technical Staff, New Grad ($140K–$175K, August 17). The posts say Ambral "helps enterprises own the intelligence behind their most important workflows" and "We graduated from YC S2025, raised millions in funding, and are already deployed within multi-billion dollar enterprises.", giving enterprises open-weight models they own and control rather than rent.Source
• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source

Create a YC-only offer: a free trace review of one workload with a written cost-and-quality report, and a discounted first quarter. Distribute it through YC's founder channels and the batch network, with a short page explaining what the team needs to send and what they get back.

Founders trust batchmates, decide quickly and talk to each other. A dozen YC customers create the logos and referrals a seed company needs.

• Free reviews swamp a small team

  Cap at five a month and require a minimum monthly spend.

• #2 Update the YC profile and post a new launch for the endpoint

• Y Combinator profileYC company page: tagline "AI for account management and customer success"; Summer 2025 batch, founded 2025, team size 4, New York. The only launch post is "Ambral Cortex: Enterprise-grade AI Account Managers". Search results and startup directories still describe Ambral as AI account managers for post-sale revenue teams.Source
• Your websiteBook page: "Bring a workload. We'll take it from there." A 20-minute walkthrough with an Ambral engineer; visitors bring "one agent or workflow" and its approximate request volume. Booking is a Calendly scheduler (or email); the page itself asks nothing about the workload.Source

Write one page per workload: what a good eval looks like for that task, typical graders, which kinds of models usually handle which share of traffic, risks to watch, and a real or anonymised result when available. Each page ends with "Bring this workload" and a pre-filled booking.

Teams search by the job their agent does, not by "inference optimisation". Workload pages meet them there and show Ambral already understands their task.

• Thin pages

  Publish only workloads the team has run, with concrete eval detail.

• Your websiteHomepage numbers, all labelled "Illustrative" or "sample data": cost per 1k requests at matched eval score, an Ambral-trained 27B model at $0.40 (eval 0.95) against a frontier closed model at $6.00 (eval 0.94) and an untuned open 27B model at $0.50 (eval 0.86); a traffic-share chart by model class for data-extraction, claims-intake and code-review. Example workloads shown: research-agent, code-review, claims-intake, contract-review and sales-outreach.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source

Send each customer a monthly report per workload: spend against the old baseline, eval scores before and after, share of traffic on each model class, items the team reviewed, and the next suggested change. Put the same numbers in a simple dashboard. Use the report for the renewal conversation and to suggest the next workload.

Cost savings are easy to forget once they become the new normal. A monthly number in the champion's inbox protects the account and gives them something to show their own boss.

• A bad month is visible

  Report honestly with the fix; it builds more trust than silence.

• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source
• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source

When the first workload's quality has held for 30 days, offer the second one with a faster onboarding (reusing the connection and review queue), a credit against setup and a joint plan for all workloads. Use the monthly report from idea 16 to pick which workload has the largest saving.

Expansion inside an account that already trusts the system is the cheapest revenue a small team can win, and it deepens the switching cost that comes from owned evals and models.

• Expanding before the first workload is stable

  Only offer after the quality trigger is met.

• #16 A monthly cost-and-quality report for every customer workload

• Your websiteBook page: "Bring a workload. We'll take it from there." A 20-minute walkthrough with an Ambral engineer; visitors bring "one agent or workflow" and its approximate request volume. Booking is a Calendly scheduler (or email); the page itself asks nothing about the workload.Source
• HiringThree open roles, all in New York (secondary San Francisco): Head of Research ($250K–$400K, posted September 1, 2026), Member of Technical Staff, Evals & Post-Training ($165K–$325K, September 1) and Member of Technical Staff, New Grad ($140K–$175K, August 17). The posts say Ambral "helps enterprises own the intelligence behind their most important workflows" and "We graduated from YC S2025, raised millions in funding, and are already deployed within multi-billion dollar enterprises.", giving enterprises open-weight models they own and control rather than rent.Source

Offer self-serve shadow mode: sign up, switch the base URL, and for two weeks Ambral passes every request through unchanged while it builds evals and replays traffic against candidate models in the background. At the end the team gets a report: projected cost at matched quality by workload, the evals used, items needing review, and a one-click pilot. Cap volume and require the Data Use terms from idea 1.

It turns the homepage promise into the buyer's own number without asking them to risk production quality, and lets a four-person team serve many more prospects than its calendar allows.

• Replay compute costs with no revenue

  Sample traffic, cap volume and require a minimum spend.

• Latency added to production calls

  Replay asynchronously; publish the pass-through latency.

• #1 Rewrite the privacy policy and terms for the inference endpoint: today they describe the Cortex account-management product
• #4 Publish a quickstart for the one-line switch, and link docs from the homepage

• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source
• Your websiteBook page: "Bring a workload. We'll take it from there." A 20-minute walkthrough with an Ambral engineer; visitors bring "one agent or workflow" and its approximate request volume. Booking is a Calendly scheduler (or email); the page itself asks nothing about the workload.Source

Turn ownership into a contract term and a feature: the customer owns weights trained on their data (within each base model's license), can export them at any time, and can choose to run them in their own cloud account. Publish it on an "Ownership" page with the license rules per base model, and add it to the enterprise path and the terms from idea 1.

Lock-in is the strongest objection to routing production traffic through a startup. A written exit path lowers the risk of saying yes, and it sets Ambral apart from routers that never produce an asset the customer keeps.

• Customers export and leave

  Keep the continuous-learning loop as the reason to stay.

• License limits vary by base model

  Show the rights per model on the Ownership page.

• #1 Rewrite the privacy policy and terms for the inference endpoint: today they describe the Cortex account-management product

• HiringThree open roles, all in New York (secondary San Francisco): Head of Research ($250K–$400K, posted September 1, 2026), Member of Technical Staff, Evals & Post-Training ($165K–$325K, September 1) and Member of Technical Staff, New Grad ($140K–$175K, August 17). The posts say Ambral "helps enterprises own the intelligence behind their most important workflows" and "We graduated from YC S2025, raised millions in funding, and are already deployed within multi-billion dollar enterprises.", giving enterprises open-weight models they own and control rather than rent.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source
• Your websiteModels page: one price table "as of November 5, 2026", in USD per 1M tokens (input, cached input, output), for 16 entries: seven open-weight models available for "evaluation, inference, and training, subject to license" and nine closed-model entries from two providers for "evaluation and inference only, no training", plus cache-write billing notes. There are no per-model pages, comparisons or change history.Source

Each quarter, publish an aggregated, consented index: for each task type (extraction, intake, code review, research and others), how often a smaller or trained model matched the frontier baseline, the typical cost ratio and where it failed. Never show any single customer's data. Release it with a methods note, a short talk and the underlying task definitions.

It makes Ambral the reference for the question every AI team asks ("can we use a cheaper model for this?"), earns press and links each quarter, and every chart points back to "test it on your traces".

• Customer data could be inferred

  Minimum group sizes, consent and review before release.

• Too few workloads early on

  Start with public tasks run through the same pipeline.

• #12 Publish the eval method behind "matched eval score" as the Head of Research's first public work

• Your websiteHomepage numbers, all labelled "Illustrative" or "sample data": cost per 1k requests at matched eval score, an Ambral-trained 27B model at $0.40 (eval 0.95) against a frontier closed model at $6.00 (eval 0.94) and an untuned open 27B model at $0.50 (eval 0.86); a traffic-share chart by model class for data-extraction, claims-intake and code-review. Example workloads shown: research-agent, code-review, claims-intake, contract-review and sales-outreach.Source
• Competitor researchRivals' own homepages: Not Diamond, "model routing for coding agents" for "frontier quality at a fraction of the price", with a savings calculator, public docs and 15 customer logos; OpenRouter, one OpenAI-compatible API for "hundreds of models" with pass-through usage pricing, a free tier and a published study of "more than 100 trillion tokens"; Fireworks, "Own your specialized intelligence", to run and train open models, self-serve and "trusted by 10,000+ companies".See the research Source
• HiringRole detail: the New Grad engineer will be "Forward deploying with our customers to understand their tasks and data", "Designing graders to turn ambiguous business objectives into verifiable rewards" and "Building replay and observability systems that make agent behavior explainable and measurable"; the Evals & Post-Training engineer is "Creating eval sets that are representative, reproducible, and resistant to overfitting" and "Scaling from individual environments to thousands of concurrent training and evaluation runs"; the Head of Research will "own the research agenda" and "help establish the research culture at Ambral Labs".Source

Pick revenue-agent builders (sales outreach, account management, renewals and customer success) as the first named vertical: a workload page with graders based on replies, meetings and renewals, a short write-up of what the team learned building Cortex, and targeted outreach to companies building agents in this space. Reuse Cortex-era relationships where they fit.

A focused first vertical gives a small team repeatable proof and shared language, and Ambral's history becomes an advantage instead of an awkward pivot.

• Looks like a step back to the old product

  Frame it as infrastructure for others' revenue agents, not Cortex again.

• #15 A landing page for each example workload on the homepage

• Y Combinator profileYC company page: tagline "AI for account management and customer success"; Summer 2025 batch, founded 2025, team size 4, New York. The only launch post is "Ambral Cortex: Enterprise-grade AI Account Managers". Search results and startup directories still describe Ambral as AI account managers for post-sale revenue teams.Source
• Your websiteHomepage numbers, all labelled "Illustrative" or "sample data": cost per 1k requests at matched eval score, an Ambral-trained 27B model at $0.40 (eval 0.95) against a frontier closed model at $6.00 (eval 0.94) and an untuned open 27B model at $0.50 (eval 0.86); a traffic-share chart by model class for data-extraction, claims-intake and code-review. Example workloads shown: research-agent, code-review, claims-intake, contract-review and sales-outreach.Source
• Your websiteHomepage (www.ambral.com redirects to labs.ambral.com): H1 "Inference that learns from your workloads.", subhead "Lower AI spend without compromising quality.", aimed "For AI-native companies"; "Change one line. Keep everything else." with a code sample that swaps one base URL; four steps, Connect, Evaluate, Optimize and Learn, ending "Accepted outputs, merged changes and closed deals flow back." The only proof is "Backed by Y Combinator": no customer logos, testimonials or case studies, no docs link, no pricing, and no mention of security, data retention or deployment options. Every call to action is "Chat with us".Source

Set a performance budget before the site grows: remove or lazy-load unused JavaScript from the homepage, load the chart components only when they scroll into view, and run a PageSpeed check in the deploy pipeline on the homepage, /models and the calculator. Keep the real-visitor numbers as the target, not the lab score.

A developer audience notices a sluggish page, and a budget set now is cheaper than a clean-up later.

• Time taken from product work

  Cap at three days; the budget prevents repeat work.

• Google PageSpeedHomepage lab test (November 2026): performance 69 on mobile and 60 on desktop; total blocking time 1,590 ms on mobile and 6,658 ms on desktop; 385 KB of JavaScript in a 523 KB page, 123 KB of it unused. Real visitors (Chrome UX Report, 75th percentile) pass Core Web Vitals: LCP 649 ms, INP 96 ms, CLS 0, with LCP 21% faster than before.Source