Astraea growth playbook

Add a short proof block to /platform/sdtm, /landing/sdtm-automation and /ads/sdtm-automation: 94% of rules correct, 100% of rows with a verified citation, 2,137 rules across 116 datasets, plus a one-line note on the method and its limits (synthetic packages, model-judged). Add a "Request an SDTM demo" button to /platform/sdtm at the top and after the workflow, and change the specbench link text to the result itself. Use the same block in demo follow-up emails and pilot proposals.

Biometrics buyers are statisticians and programmers who trust measured numbers over adjectives. Every vendor in the category says "AI SDTM automation". Few show a rule-level accuracy figure with its method. Putting it next to the call to action turns curiosity into a demo request.

Measured, not claimed
94% of SDTM mapping rules judged correct across 116 datasets and 2,137 rules, against 84-87% for general-purpose coding agents.
100% of specification rows cite a verified source quotation.
Method: six synthetic study packages from specbench, judged against known references. Your study still needs your own validation and review.
[Request an SDTM demo]  [See the method]

• Buyers read a synthetic benchmark as a guarantee for their study

  State the method and limits next to the number, as the post already does, and offer a pilot on a study with known outputs.

• Naming rival tools invites disputes

  Describe them as general-purpose coding agents on the product pages and leave detail to the post.

• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteSDTM product page: title "Astraea SDTM | AI Mapping, Specifications & Datasets"; a five-step workflow and its deliverables (datasets, editable specs, programs, Define-XML, annotated CRF, reviewer's guide); no performance figures; none of its links goes to the demo form, the main actions are "Read Astraea's technical overview" and "Explore ADaM", and specbench appears only as a link, "How we measure SDTM spec accuracy".Source
• Your websitePaid landing page /ads/sdtm-automation (listed in the sitemap): H1 "SDTM mapping software. Automation you can inspect."; CTA "Request an SDTM demo"; its own form (first name, last name, work email, optional organization and "What would you like to automate?") with a note not to send patient data; no customer proof or performance figures; "Agency acceptance cannot be guaranteed."Source
• Your websiteLanding page /landing/sdtm-automation: "Source mapping is a study decision, not just a conversion."; suggests starting with one representative domain before wider deployment; CTA "Book a working session"; no quantitative claims, logos or case studies.Source

Update the row to cite specbench (94% of rules correct, 100% cited rows) and the launch-post timing, label what each benchmark measures (synthetic packages vs one completed study), and add a "last checked" date to every compare page. Check the other three comparisons for rows that the October posts have made out of date.

Comparison pages are read by buyers who are actively choosing. A row your own page marks against you is the line they remember.

• Benchmarks are not like for like

  Say plainly what each measures rather than claiming a win.

• Your websiteComparison page /compare/koneksa-lodestar: the row "Vendor-published study metrics" scores Astraea "Partial" because it has no public single-study benchmark, while quoting Koneksa's results from one study (80% fewer programming hours, 65% lower biometrics cost, 16x faster QC); it also marks eCTD, BIMO and EMA submission as not in Astraea's listed scope.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteBlog, "Introducing Astraea" (October 19, 2026): internal development time fell from "roughly two weeks" to "just under five hours"; a chart compares an 8-week manual biometrics workflow (from a Takeda case study at PHUSE US Connect 2023) with 5 hours in Astraea; it mentions first enterprise deployments but names no customer.Source

Use one form on every page. Required: name and work email. Optional: phone. Add short questions the team needs to scope a pilot: workflow (SDTM, ADaM, TFL, data management, visualization), phase, therapeutic area, R or SAS, and preferred deployment. Show a scheduler right after submission, and a line on what happens next (a 30-minute call, then a scoping note within two working days). Route by workflow so the CTO joins technical requests.

A required phone field cuts form completion for enterprise buyers who prefer email. Scoping questions mean the first call opens with a pilot proposal instead of a discovery interview.

• More questions lower completion

  Keep scoping questions optional and one click each.

• Your websiteDemo page /book-a-demo: "Have a question or want to see Astraea?"; three required fields (name, business email, phone); no scheduler and no next step beyond "We'll use your details to respond to this request"; no proof on the page. The site removed its Cal.com scheduler after April 2026.Source
• Your websitePaid landing page /ads/sdtm-automation (listed in the sitemap): H1 "SDTM mapping software. Automation you can inspect."; CTA "Request an SDTM demo"; its own form (first name, last name, work email, optional organization and "What would you like to automate?") with a note not to send patient data; no customer proof or performance figures; "Agency acceptance cannot be guaranteed."Source
• Your websiteFAQ: teams can start with one workflow, such as SDTM mapping; "Choose a representative study or workflow with known outputs", agree acceptance criteria and involve IT and QA early; what is sent to an external model depends on the approved deployment, with client-owned API keys and self-hosted options discussed; "SAS execution connects to your licensed SAS environment"; per-study and broader licences have been discussed; EDC write-back is a separate integration.Source

Decide what each number means: 5 hours of system run time for one internal study, 2 days of elapsed time including expert review. State both with their basis, for example "About 5 hours of generation and 2 days to reviewed outputs, against 8 weeks manually", and use the same wording on the homepage, the SDTM pages, the YC profile and sales decks.

Precise, consistent claims build trust with technical buyers, and explaining the gap between generation and review matches the site's own human-review message.

• A smaller number sounds less impressive

  Lead with the 8-week comparison, which is the real gain.

• Your websiteHomepage: H1 "SDTM, ADaM, and TFL in 2 days" over "Post-protocol work through submission preparation, inside your environment."; three modules (Data Management, Data Analysis, Data Visualization); no customer logos or testimonials; the FAQ says pricing is scoped around modules, study workload, deployment and support; a Gold Sponsor of BBSW 2026 banner runs on every page; the only profiles linked are Y Combinator and LinkedIn.Source
• Your websiteBlog, "Introducing Astraea" (October 19, 2026): internal development time fell from "roughly two weeks" to "just under five hours"; a chart compares an 8-week manual biometrics workflow (from a Takeda case study at PHUSE US Connect 2023) with 5 hours in Astraea; it mentions first enterprise deployments but names no customer.Source
• Company profileY Combinator profile: Spring 2026 batch, founded 2026, team of 6, San Francisco; automates clinical trial biometrics "from raw study inputs to FDA-ready outputs" (data normalization, CDISC mapping, SDTM, ADaM, TFLs, QC and revisions), aimed at Phase II/III sponsors.Source

Audit every /solutions/ and /platform/ page for a visible next step. Add a call to action that matches the page: "Scope a one-study pilot" on biotech and Phase 1-2, "Plan an enterprise evaluation" on enterprise pharma, "Request an SDTM demo" on SDTM. Each goes to the single demo form with the workflow pre-filled.

Buyers who reach a solutions page are segmenting themselves. A call to action in their own terms converts better than a generic "Request a demo", or none.

• Too many buttons feel pushy for technical readers

  Two per page, with plain wording.

• #3 One demo form everywhere: ask about the study and workflow instead of a phone number, and let buyers pick a time

• Your websiteBiotech solutions page: written for small biotech teams with limited capacity; no demo or contact call to action (its links go to the technical overview, the Phase 1-2 page, the technical FAQ and "What is clinical trial automation?"); no metrics or case studies.Source
• Your websiteSitemap (58 URLs, checked November 9, 2026, no lastmod dates): three pages aimed at SDTM automation (/platform/sdtm, /landing/sdtm-automation, /ads/sdtm-automation) plus /landing/sdtm-qc and /landing/cdisc-dataset-automation; 20 /landing/ pages, also listed on a /landing-pages page titled "Search landing pages"; four /compare pages; solutions pages for biotech, enterprise pharma, embedded teams, Phase 1-2 and Phase 3; a /landing/clinical-trial-reprogramming page; four blog posts; no case-study or customer pages.Source

Pick /platform/sdtm as the page that should rank for SDTM automation. Mark /ads/sdtm-automation noindex and drop it from the sitemap (it is for paid traffic). Point the canonical of /landing/sdtm-automation to /platform/sdtm, or rewrite it for a distinct query. Drop /landing-pages from the sitemap and noindex it. Add lastmod dates to the sitemap. Then link /platform/sdtm from the landing pages so they pass authority to it.

One strong page per query ranks better than three similar ones, and keeping ad pages out of search keeps paid tests clean.

• A landing page already ranks

  Check Search Console first and keep whichever page ranks best as the main one.

• Your websiteSitemap (58 URLs, checked November 9, 2026, no lastmod dates): three pages aimed at SDTM automation (/platform/sdtm, /landing/sdtm-automation, /ads/sdtm-automation) plus /landing/sdtm-qc and /landing/cdisc-dataset-automation; 20 /landing/ pages, also listed on a /landing-pages page titled "Search landing pages"; four /compare pages; solutions pages for biotech, enterprise pharma, embedded teams, Phase 1-2 and Phase 3; a /landing/clinical-trial-reprogramming page; four blog posts; no case-study or customer pages.Source
• Your websitePaid landing page /ads/sdtm-automation (listed in the sitemap): H1 "SDTM mapping software. Automation you can inspect."; CTA "Request an SDTM demo"; its own form (first name, last name, work email, optional organization and "What would you like to automate?") with a note not to send patient data; no customer proof or performance figures; "Agency acceptance cannot be guaranteed."Source
• Your websiteLanding page /landing/sdtm-automation: "Source mapping is a study decision, not just a conversion."; suggests starting with one representative domain before wider deployment; CTA "Book a working session"; no quantitative claims, logos or case studies.Source

Review the tags loaded through the tag manager and remove any that are unused or duplicated. Load analytics after the main content. Check the Astro islands for scripts shipped to pages that don't need them. Enlarge the small touch targets. Re-run PageSpeed and keep the report with each release.

A small, cheap cleanup that keeps mobile visits fast and tidies tool ownership as the GTM stack grows.

• Removing a tag breaks reporting

  Confirm the owner of each tag before removing it.

• Speed testGoogle PageSpeed (November 9, 2026): mobile score 82, desktop 100; in the lab the mobile main content appears after 3.3s, but real visitors pass Core Web Vitals (75th-percentile LCP 1.2s, INP 65ms, CLS 0) and LCP has improved 22%; about 314 KB of unused JavaScript, with Google Tag Manager the largest third party at 149 KB; one accessibility issue: touch targets too small or too close.Source

Name and describe a pilot package: one completed study with known outputs, one workflow (SDTM, ADaM or TFL), a fixed duration (for example six weeks once infrastructure is approved), a written acceptance-criteria template, a side-by-side comparison report at the end, and a pilot fee credited against the first licence. Publish it on a /pilot page with a downloadable one-pager and the evaluation checklist.

A fixed scope shortens procurement, sets expectations for both sides and gives the founders a repeatable sales motion while the team is six people.

[ ] Study chosen with known reference outputs
[ ] One workflow in scope (SDTM / ADaM / TFL)
[ ] Inputs listed: protocol, aCRF, SAP, raw data, prior specs
[ ] Acceptance criteria: variable match rate, open questions, review hours
[ ] Reviewers named for each output
[ ] IT and QA sign-off on deployment
[ ] End date and comparison report agreed

• Fixed scope underprices complex studies

  Limit the package to one workflow and define what triggers a custom scope.

• Your websiteFAQ: teams can start with one workflow, such as SDTM mapping; "Choose a representative study or workflow with known outputs", agree acceptance criteria and involve IT and QA early; what is sent to an external model depends on the approved deployment, with client-owned API keys and self-hosted options discussed; "SAS execution connects to your licensed SAS environment"; per-study and broader licences have been discussed; EDC write-back is a separate integration.Source
• Your websiteCompare pages: Astraea vs traditional CRO and in-house teams, vs Saama (described as a broad cloud clinical data suite with published enterprise relationships with Pfizer and AstraZeneca), vs TriClick and Hill Research (cloud software with self-serve plans metered by per-table credits, plus biostatistics FSP services) and vs Koneksa Lodestar (early access).Source
• Your websiteBiotech solutions page: written for small biotech teams with limited capacity; no demo or contact call to action (its links go to the technical overview, the Phase 1-2 page, the technical FAQ and "What is clinical trial automation?"); no metrics or case studies.Source

Build a downloadable review pack (gated by a short request form): a data-flow diagram for each deployment option (air-gapped, self-hosted model, client-owned keys) showing what, if anything, leaves the environment; a Part 11 controls table (audit trail, access, e-signature) with how each is met; a validation starter kit (intended-use statement, test scripts the client can run); SOC 2 status and the planned scope, stated factually; and a pre-filled standard security questionnaire.

In regulated software, deals stall in review, not in the demo. Handing IT and QA their answers on day one shortens the cycle and makes a small vendor look prepared.

• Overstating compliance

  State what is done, in progress and client-owned, as the security page already does.

• Your websiteSecurity page (titled Trust Center): "SOC 2 is in progress" and Astraea "has not completed a SOC 2 examination"; HIPAA-aligned safeguards; built for "21 CFR Part 11-aligned workflows"; air-gapped deployment available; no downloadable reports or policy documents.Source
• Your websiteAbout page: "Astraea is a software platform, not a contract research organization."; forward-deployed engineers install it in each sponsor's environment; "Astraea does not see customer data."; the company started in oncology and rare disease; backed by Y Combinator.Source
• Your websiteFAQ: teams can start with one workflow, such as SDTM mapping; "Choose a representative study or workflow with known outputs", agree acceptance criteria and involve IT and QA early; what is sent to an external model depends on the approved deployment, with client-owned API keys and self-hosted options discussed; "SAS execution connects to your licensed SAS environment"; per-study and broader licences have been discussed; EDC write-back is a separate integration.Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source

Agree with one deployment customer an anonymised case study: therapeutic area and phase, the workflow, inputs, time from inputs to reviewed outputs against their usual baseline, the share of rules reviewers accepted unchanged, and one quote from a named role (not a person). Publish it as /customers/<slug>, add it to the SDTM and solutions pages, and use it in every proposal.

Peer evidence from a real study answers the question a benchmark cannot: did it work on someone else's messy data?

• The customer declines

  Offer a metrics-only summary with no identifying details.

• Your websiteBlog, "Introducing Astraea" (October 19, 2026): internal development time fell from "roughly two weeks" to "just under five hours"; a chart compares an 8-week manual biometrics workflow (from a Takeda case study at PHUSE US Connect 2023) with 5 hours in Astraea; it mentions first enterprise deployments but names no customer.Source
• Your websiteSitemap (58 URLs, checked November 9, 2026, no lastmod dates): three pages aimed at SDTM automation (/platform/sdtm, /landing/sdtm-automation, /ads/sdtm-automation) plus /landing/sdtm-qc and /landing/cdisc-dataset-automation; 20 /landing/ pages, also listed on a /landing-pages page titled "Search landing pages"; four /compare pages; solutions pages for biotech, enterprise pharma, embedded teams, Phase 1-2 and Phase 3; a /landing/clinical-trial-reprogramming page; four blog posts; no case-study or customer pages.Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source

Each month, pull registered Phase II/III interventional studies in oncology and rare disease with a primary completion date in the next two to six months, sponsored by small and mid-size companies. Find the head of biometrics or statistical programming at each, and send a short note tied to that study's milestone, offering the fixed pilot on a completed study. Log every account in the CRM with the trigger date.

Outreach timed to a real milestone is relevant, not cold, and the same list tells the team which therapeutic areas to build examples for.

• Work is already committed to a CRO

  Pitch the next study or a reprogramming project instead.

• #7 Publish the pilot as a fixed package: one study, one workflow, known outputs, agreed acceptance criteria

• Company profileY Combinator profile: Spring 2026 batch, founded 2026, team of 6, San Francisco; automates clinical trial biometrics "from raw study inputs to FDA-ready outputs" (data normalization, CDISC mapping, SDTM, ADaM, TFLs, QC and revisions), aimed at Phase II/III sponsors.Source
• Your websiteAbout page: "Astraea is a software platform, not a contract research organization."; forward-deployed engineers install it in each sponsor's environment; "Astraea does not see customer data."; the company started in oncology and rare disease; backed by Y Combinator.Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source

Write a repeatable playbook for each biometrics event. Two weeks before, research attendees and pre-book meetings. At the booth, show a two-minute specbench walk-through and run a live SDTM example on public data. Scan every conversation into the CRM with a workflow tag. Within 48 hours, send a follow-up offering the fixed pilot. Afterwards, measure pipeline per event to decide which events to sponsor next.

Biometrics is a small, conference-driven community. Following up with specifics turns a booth visit into a scoped call.

• Booth demos fail on conference Wi-Fi

  Run the example offline on a laptop.

• #7 Publish the pilot as a fixed package: one study, one workflow, known outputs, agreed acceptance criteria

• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source
• Your websiteHomepage: H1 "SDTM, ADaM, and TFL in 2 days" over "Post-protocol work through submission preparation, inside your environment."; three modules (Data Management, Data Analysis, Data Visualization); no customer logos or testimonials; the FAQ says pricing is scoped around modules, study workload, deployment and support; a Gold Sponsor of BBSW 2026 banner runs on every page; the only profiles linked are Y Combinator and LinkedIn.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source

Add an RSS feed (Astro supports it out of the box) and a one-field subscribe box at the end of every post and on the specbench post. Send a short monthly note: one new post, one specbench or product update, and one practical tip for statistical programmers. Score subscribers who open three notes as warm leads.

It keeps Astraea in front of statisticians through a cycle that takes months, at almost no cost.

• Not enough content to send

  Monthly is enough, and product changes count as content.

• Your websiteBlog: four posts (July 4, October 5, October 19 and October 27, 2026); no RSS feed, newsletter or subscribe option.Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source

Publish one guide per common SDTM domain (DM, AE, EX, MH, CM, LB, VS, SUPPQUAL): the core variables, common source patterns, the mistakes specbench found, and a worked example on public synthetic data, ending with an invitation to the pilot. Start with the domains where specbench showed errors.

Practitioner content earns links and repeat visits from the people who later champion a tool, and it shows expertise better than a product page can.

• Errors in public guides hurt credibility

  Technical review before every guide is published.

• Your websiteSitemap (58 URLs, checked November 9, 2026, no lastmod dates): three pages aimed at SDTM automation (/platform/sdtm, /landing/sdtm-automation, /ads/sdtm-automation) plus /landing/sdtm-qc and /landing/cdisc-dataset-automation; 20 /landing/ pages, also listed on a /landing-pages page titled "Search landing pages"; four /compare pages; solutions pages for biotech, enterprise pharma, embedded teams, Phase 1-2 and Phase 3; a /landing/clinical-trial-reprogramming page; four blog posts; no case-study or customer pages.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiterobots.txt: lets AI search and answer crawlers in (Content-Signal search=yes, ai-input=yes) and blocks training crawlers (ai-train=no).Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source

Design a partner programme for two or three mid-size CROs or FSP providers: a licence priced per study delivered, training and certification for their programmers, joint pilots with their sponsors, and co-branded case studies. Keep sponsor-direct deals for companies that bring the work in house.

One CRO partner gives access to dozens of sponsor studies without a separate enterprise sale for each, and it changes CROs from rivals into a channel.

• Channel conflict with direct deals

  Register deals and pay partners only for studies they bring.

• Compare pages offend partners

  Reframe the CRO comparison around operating models, not replacement.

• #7 Publish the pilot as a fixed package: one study, one workflow, known outputs, agreed acceptance criteria

• Your websiteAbout page: "Astraea is a software platform, not a contract research organization."; forward-deployed engineers install it in each sponsor's environment; "Astraea does not see customer data."; the company started in oncology and rare disease; backed by Y Combinator.Source
• Your websiteCompare pages: Astraea vs traditional CRO and in-house teams, vs Saama (described as a broad cloud clinical data suite with published enterprise relationships with Pfizer and AstraZeneca), vs TriClick and Hill Research (cloud software with self-serve plans metered by per-table credits, plus biostatistics FSP services) and vs Koneksa Lodestar (early access).Source
• HiringFounding GTM Lead (San Francisco, open on Y Combinator): prepare for and follow up on conferences such as the Bay Area Biotech-Pharma Statistics Workshop (BBSW); run enterprise sales end to end with biometrics, programming, IT, security and procurement stakeholders; build outbound campaigns and partnerships and own the CRM; write positioning, sales materials, website copy and case studies; develop inbound demand through content, webinars and SEO experiments; measure the funnel and automate research, outreach and reporting.Source
• Company profileY Combinator profile: Spring 2026 batch, founded 2026, team of 6, San Francisco; automates clinical trial biometrics "from raw study inputs to FDA-ready outputs" (data normalization, CDISC mapping, SDTM, ADaM, TFLs, QC and revisions), aimed at Phase II/III sponsors.Source

Wire the full specbench set into the release pipeline: every release runs all 40 packages, records rule accuracy, coverage and citation rates, and blocks the release if accuracy drops beyond a threshold. Publish a simple accuracy-by-release chart on the specbench page and include the release record in the validation kit for clients.

It catches regressions before customers do, gives QA reviewers at sponsors evidence of controlled change, and turns a one-off blog post into a living proof point.

• Overfitting to the benchmark

  Keep a held-out set of packages that is never used in development.

• HiringQA Engineer (remote contract, open on Y Combinator): manual, exploratory, smoke and regression testing of core workflows, including error states, permissions and data inputs; maintain a regression checklist; clinical trials experience is a plus.Source
• HiringFounding Engineer (San Francisco, open on Y Combinator, $150K-$300K): build agentic workflows for clinical trial automation; design reliable systems for document ingestion, data transformation, validation and auditability; integrate with legacy systems; ship fast with the rigor regulated healthcare needs.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteBlog, "Introducing Astraea" (October 19, 2026): internal development time fell from "roughly two weeks" to "just under five hours"; a chart compares an 8-week manual biometrics workflow (from a Takeda case study at PHUSE US Connect 2023) with 5 hours in Astraea; it mentions first enterprise deployments but names no customer.Source

Make a standard deliverable for the first week after the environment is ready: one SDTM domain generated from the sponsor's own inputs, with each variable's source citation, the rules flagged for decisions and a comparison with the known reference. Hold a 30-minute review meeting on day five and record the time reviewers needed to check it. Use the same format for every pilot.

Early, concrete output keeps sponsor attention through the slow parts of deployment and gives the champion something to show internally.

• Inputs are incomplete in week one

  Choose the domain with the cleanest inputs and flag the gaps, as the FAQ promises.

• #7 Publish the pilot as a fixed package: one study, one workflow, known outputs, agreed acceptance criteria

• Your websiteLanding page /landing/sdtm-automation: "Source mapping is a study decision, not just a conversion."; suggests starting with one representative domain before wider deployment; CTA "Book a working session"; no quantitative claims, logos or case studies.Source
• Your websiteFAQ: teams can start with one workflow, such as SDTM mapping; "Choose a representative study or workflow with known outputs", agree acceptance criteria and involve IT and QA early; what is sent to an external model depends on the approved deployment, with client-owned API keys and self-hosted options discussed; "SAS execution connects to your licensed SAS environment"; per-study and broader licences have been discussed; EDC write-back is a separate integration.Source
• Your websiteBlog, "The next bottleneck in clinical development is evidence continuity" (October 5, 2026): asks teams to measure the cost of reconstruction on three recent review questions and cites the CDISC Analysis Results Standard; no call to action.Source

Package a legacy conversion offer: bring 3-10 completed studies to one current SDTM version with traceable mappings, starting with one study as the pilot and checking against the original outputs. Write a dedicated page and a short outbound sequence for sponsors preparing pooled analyses, and use the converted studies as a stepping stone to live-study work.

Completed studies carry little risk, the success test is built in, and conversion projects are tedious enough that sponsors are glad to hand them over.

• Old studies have poor source documents

  Score input quality before quoting.

• #7 Publish the pilot as a fixed package: one study, one workflow, known outputs, agreed acceptance criteria

• Your websiteSitemap (58 URLs, checked November 9, 2026, no lastmod dates): three pages aimed at SDTM automation (/platform/sdtm, /landing/sdtm-automation, /ads/sdtm-automation) plus /landing/sdtm-qc and /landing/cdisc-dataset-automation; 20 /landing/ pages, also listed on a /landing-pages page titled "Search landing pages"; four /compare pages; solutions pages for biotech, enterprise pharma, embedded teams, Phase 1-2 and Phase 3; a /landing/clinical-trial-reprogramming page; four blog posts; no case-study or customer pages.Source
• Your websiteFAQ: teams can start with one workflow, such as SDTM mapping; "Choose a representative study or workflow with known outputs", agree acceptance criteria and involve IT and QA early; what is sent to an external model depends on the approved deployment, with client-owned API keys and self-hosted options discussed; "SAS execution connects to your licensed SAS environment"; per-study and broader licences have been discussed; EDC write-back is a separate integration.Source

The CEO and CTO each post once a week: one specbench finding, one lesson from a deployment (anonymised), and one opinion on review and traceability. Each post links to a guide or the pilot page. Repost the best ones as the company. Track profile visits and demo requests that mention LinkedIn.

In a niche field, steady, specific posts from builders become the reason a statistician recognises the name at a conference or in a vendor shortlist.

• Posting drifts when the founders are busy

  Keep a four-week backlog drafted ahead.

• Your websiteHomepage: H1 "SDTM, ADaM, and TFL in 2 days" over "Post-protocol work through submission preparation, inside your environment."; three modules (Data Management, Data Analysis, Data Visualization); no customer logos or testimonials; the FAQ says pricing is scoped around modules, study workload, deployment and support; a Gold Sponsor of BBSW 2026 banner runs on every page; the only profiles linked are Y Combinator and LinkedIn.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteBlog, "Introducing Astraea" (October 19, 2026): internal development time fell from "roughly two weeks" to "just under five hours"; a chart compares an 8-week manual biometrics workflow (from a Takeda case study at PHUSE US Connect 2023) with 5 hours in Astraea; it mentions first enterprise deployments but names no customer.Source

Release the synthetic packages, the reference specifications and the scoring code under an open licence, keeping a held-out set private to prevent gaming. Host a public leaderboard, invite other vendors, CROs and academic groups to submit, and present the benchmark at statistical programming conferences. Report Astraea's own score on every release.

The company that defines the measure is quoted whenever it is used, and biometrics teams trust a benchmark they can run themselves.

• A competitor scores higher

  That is still a win for the benchmark's authority; publish the improvement roadmap.

• Model-judged scoring is challenged

  Publish the code checks separately and invite human spot-checks.

• #15 Run specbench on every release as a regression suite, and publish the accuracy trend

• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteComparison page /compare/koneksa-lodestar: the row "Vendor-published study metrics" scores Astraea "Partial" because it has no public single-study benchmark, while quoting Koneksa's results from one study (80% fewer programming hours, 65% lower biometrics cost, 16x faster QC); it also marks eCTD, BIMO and EMA submission as not in Astraea's listed scope.Source
• Competitor researchSaama's homepage: "Accelerate Your Trials Intelligently"; SaaS products plus industry services; claims "Up to 62% less programming effort" for biostatistics and submissions, with BRAIN SDTM in its statistical programming suite.See the research Source

Build a hosted sandbox limited to synthetic study packages: a programmer picks a study, asks for one SDTM domain or a chart in plain language, and sees the specification, the R code, the output and the source citations. Require a work email after the first result, and pass the session to sales with what was tried.

Statistical programmers decide by inspecting code, and a sandbox lets champions test the product before procurement starts, without any study data involved.

• Visitors upload real study data

  Allow only the provided synthetic packages, with no uploads.

• Sandbox results disappoint

  Limit it to tasks the product handles well and say so.

• Your websiteData Analysis and Data Visualization pages: static examples only (the R and SAS panels read "Code will appear here"; the Kaplan-Meier chart uses illustrative data); "Request a demo" is the only action, with no sandbox or downloadable sample output.Source
• Your websiteBlog, "Measuring SDTM spec-writing accuracy with specbench" (October 27, 2026): on six synthetic study packages (116 datasets, 2,137 inferable mapping rules), 94% of Astraea's rules were judged correct, against 87% and 84% for two general-purpose coding agents, and 100% of Astraea's rows carried a verified quotation (74% and 65% for the others); the full benchmark has 40 packages and 15,709 rules; the post links no public specbench repository and notes the packages are synthetic and accuracy was judged by a model.Source
• Your websiteAbout page: "Astraea is a software platform, not a contract research organization."; forward-deployed engineers install it in each sponsor's environment; "Astraea does not see customer data."; the company started in oncology and rare disease; backed by Y Combinator.Source

Scope the submission layer in stages: first, generate and check the reviewer's guides (SDRG and ADRG) and BIMO listings from the traced data already in Astraea; then assemble the Module 5.3 datasets folder structure for the sponsor's publishing tool. Leave eCTD publishing to existing tools and integrate rather than replace. Update the comparison row as each stage ships.

Traceability already built into the data makes the reviewer's guides and BIMO listings a natural extension, and it removes a row where Astraea now loses.

• Scope creep for a team of six

  Ship reviewer's guides first and only build further where customers pay.

• Your websiteComparison page /compare/koneksa-lodestar: the row "Vendor-published study metrics" scores Astraea "Partial" because it has no public single-study benchmark, while quoting Koneksa's results from one study (80% fewer programming hours, 65% lower biometrics cost, 16x faster QC); it also marks eCTD, BIMO and EMA submission as not in Astraea's listed scope.Source
• Your websiteSDTM product page: title "Astraea SDTM | AI Mapping, Specifications & Datasets"; a five-step workflow and its deliverables (datasets, editable specs, programs, Define-XML, annotated CRF, reviewer's guide); no performance figures; none of its links goes to the demo form, the main actions are "Read Astraea's technical overview" and "Explore ADaM", and specbench appears only as a link, "How we measure SDTM spec accuracy".Source
• Your websiteHomepage: H1 "SDTM, ADaM, and TFL in 2 days" over "Post-protocol work through submission preparation, inside your environment."; three modules (Data Management, Data Analysis, Data Visualization); no customer logos or testimonials; the FAQ says pricing is scoped around modules, study workload, deployment and support; a Gold Sponsor of BBSW 2026 banner runs on every page; the only profiles linked are Y Combinator and LinkedIn.Source

Emit ARS-structured results metadata alongside every generated TFL: the analysis, population, method, data snapshot and program version behind each displayed number. Add a reverse view in which a reviewer clicks a number to see its contributing records and derivation rules. Publish a short technical note and an example on synthetic data.

It turns a differentiating idea into something reviewers can use, and it puts Astraea early on a standard that sponsors will meet anyway.

• Sponsors are not using ARS yet

  Sell the reverse view first; the standard is the format underneath.

• Your websiteBlog, "The next bottleneck in clinical development is evidence continuity" (October 5, 2026): asks teams to measure the cost of reconstruction on three recent review questions and cites the CDISC Analysis Results Standard; no call to action.Source
• Your websiteHomepage: H1 "SDTM, ADaM, and TFL in 2 days" over "Post-protocol work through submission preparation, inside your environment."; three modules (Data Management, Data Analysis, Data Visualization); no customer logos or testimonials; the FAQ says pricing is scoped around modules, study workload, deployment and support; a Gold Sponsor of BBSW 2026 banner runs on every page; the only profiles linked are Y Combinator and LinkedIn.Source