# Harbor

Aiolum story: https://aiolum.com/story/harbor/ · added 2026-11-05 · updated 2026-11-06

**In short.** Harbor (AI electronic data capture software for clinical trials, USA): US$15.6K a month in software revenue, May 2026. Evidence grade D: reported, not verified. Team: 2 co-founders; 4 people listed by Y Combinator in November 2026. Launched June 2025; still running.

AI electronic data capture software for clinical trials · USA · AI-native · team of 2 to 10 · still running · group: Clinic and practice admin

- Outcome: US$15.6K a month in software revenue, May 2026 (evidence grade of the headline number: D; headline in US dollars: 187200)
- Launch: June 2025; months from launch to the outcome: 11
- Team: 2 co-founders; 4 people listed by Y Combinator in November 2026
- Outside funding: Y Combinator, Spring 2026 batch; amount not published in the sources read
- Main factor: Years inside the work: the CEO helped run the clinical trials of a medical device startup and saw site staff type paper records into the database by hand
- Meters, 1 to 4: evidence strength 1, copyable pattern 1, capital needed 2
- Tags: AI-native, Small team, Funded, Running, B2B, Health Care

Links:

- Visit runharbor.com: https://runharbor.com/

**The product.** Software that drug and medical device companies use to collect the data of a clinical trial. It reads the study protocol and drafts the data entry forms, reads scanned paper or electronic records from the clinics and fills the forms for a person to confirm, and checks the data for problems. Harbor's pricing page says commercial studies start at US$2,000 a month and university studies are free.

**Why it worked.** The founders say that setting up this kind of database by hand takes companies two to three months per trial, and that clinic staff spend hours typing records into it. Harbor turns the protocol into a draft database in about 10 minutes and the paper records into form entries with a confidence score for each value, so a small sponsor can start a study in days.

Evidence grades: A = disclosed by a public company, a filing or a payment processor, or verified by one; B = two or more independent reports agree, or a journalist confirmed it; C = a founder claim that no one else verified; D = reported, not verified: one founder claim or a third-party estimate with no second source; E = no revenue figure found, so the headline is a non-money fact and the other facts carry their own grades; calc = computed on this page from graded facts.

## Summary

Harbor makes an EDC (electronic data capture system), the database in which a clinical trial, a study that tests a drug or a medical device on people, records what happens to each study subject. The company that pays for the trial is the sponsor; the clinics that see the subjects are the sites. Harbor's "Magic Build" reads the trial plan, the protocol, and drafts the CRFs (case report forms, the screens for each visit). "Magic Capture" reads scanned paper or electronic records and fills the forms, with a confidence score for each value, and a person confirms every value. The system is built to support 21 CFR Part 11, the US Food and Drug Administration (FDA) rule for electronic records and electronic signatures, so the data can go into an application to the FDA.

Albert Cai, who had helped run clinical trials and regulatory strategy at the medical device startup Biolinq, and Nathan Leung, a former engineer at Google and Ramp, started the company in about June 2025. By March 2026 it ran in five studies. In the founders' second Y Combinator launch post, on 29 May 2026, they wrote that the software had grown to seven trials and US$15,600 in monthly software revenue. Y Combinator's company page, read in November 2026, says the founders described about US$187,000 in contracted software revenue over the next twelve months.

In the same post the founders announced a turn: Harbor now also sells full-service trial work as a CRO (contract research organization, a company that runs trials for sponsors), and the Y Combinator page says its first such contract is worth US$1.93 million over three years. This page reports only the software figure, which rests on the founders alone (grade D). The Harbor website still sells the software on its own with published prices.

## Pricing

Harbor prices read on its pricing page on 5 November 2026 (grade C)

| Plan | Per month, billed yearly (US$) | Per month, billed monthly (US$) | Limit |
|---|---|---|---|
| Academic & Non-Profit | 0 | 0 | university and investigator-initiated studies; unlimited users, patients and sites |
| Commercial Study |  | from 2,000 | adds the 21 CFR Part 11 validation package, Slack support, site training, uptime SLA |
| Enterprise / CRO | volume pricing |  | 3 or more active trials; MSA, SSO, CTMS and eTMF integrations |

Commercial pricing is a starting price. The pricing page says the final price depends on study complexity, site count, the modules used (for example randomization, medical coding, ePROs) and the number of studies. Harbor says it does not charge per subject, per site or per CRF. A case study from March 2026 names a flat fee of US$2,000 for one EDC build. The free plan is for university and investigator-initiated studies; commercially sponsored or regulatory-submission studies go to the Commercial plan.

## How much revenue does Harbor make? The numbers, graded

Founder-reported software revenue of US$15,600 a month in May 2026, from seven trials, with a larger service contract signed beside it

| Fact | Value | Grade | Note | Sources |
|---|---|---|---|---|
| Monthly software revenue, May 2026 | US$15,600 | D | The founders in their Y Combinator launch post of 29 May 2026: 'grew to seven trials and $15.6k in monthly software revenue'. No second dated source. | [1] |
| Run rate from the May 2026 figure | US$187,200 a year | calc | Calculated: US$15,600 times 12. The headline amount on this page. | [1] |
| Contracted software revenue | about US$187,000 over twelve months | D | Y Combinator company page, read 5 November 2026, written by the company and not dated: '~$187k in contracted software revenue over the next twelve months'. It equals the May figure times 12, so it is not an independent check. | [2] |
| Average software revenue per trial, May 2026 | about US$2,229 a month | calc | Calculated: US$15,600 divided by the seven trials the founders named. Close to the published Commercial starting price of US$2,000 a month. | [1, 5] |
| First full-service CRO contract | US$1.93 million over three years | D | Y Combinator company page. The launch post calls it 'our first seven-figure contract' with 'a publicly traded medical device company'. A contract value, not revenue received, and service revenue, not software. | [1, 2] |
| CRO contract per year | about US$643,333 | calc | Calculated: US$1.93 million divided by 3, assuming even payments, which no source states. About 3.4 times the software run rate. | [2] |
| Studies on the software | 5 in March 2026, 7 in May 2026 | C | First launch post, 16 March 2026 ('live in five clinical studies'), and second launch post, 29 May 2026. | [3, 1] |
| Company start | about June 2025 | C | Blog post of 9 November 2025: 'Five months ago, we started Harbor Labs'. Y Combinator lists 2025 as the year founded. | [9, 2] |
| Team | 2 co-founders; 4 people | C | Y Combinator company page, read 5 November 2026: team size 4. Founders Albert Cai (CEO) and Nathan H. Leung (CTO). | [2, 6] |
| Funding | Y Combinator, Spring 2026 batch | C | Y Combinator company page. No amount appears in the sources read. | [2] |
| Location | San Francisco (Y Combinator page); the website footer says 'Made in San Diego, California' | C | Company entity on the website: Harbor OpCo, Inc. | [2, 4] |
| Time to a draft database | about 10 minutes from the protocol PDF | C | Company claim on the home and pricing pages. The time to launch depends on review, testing and validation. | [4] |
| Data entry time in the company's own test | about 2 hours by hand, 10 to 15 minutes with Magic Capture | C | Blog post of 13 October 2025: four clinic coordinators entered the same 37-page set of documents. The company's test, not an outside study. | [8] |
| Fastest customer build | 9 days from protocol to live database | C | Case study of 17 March 2026, MotilityCount ApS and NOVA Genomics, a study of about 1,600 couples. The customer said other vendors quoted 10 to 12 weeks. Flat fee for the build: US$2,000. | [10] |
| AI providers listed | Google (Vertex AI/Gemini) and OpenAI; Reducto and Extend for document extraction | C | Subprocessor registry, last updated 19 October 2026, all effective 25 May 2026. No source says which model does which task. | [7] |
| Time from start to the May 2026 figure | 11 months | calc | Calculated: June 2025 to May 2026. | [9, 1] |

How to read this table: a grade D value is one founder statement with one date and no second source. A grade C value is a statement on the company's own pages. A "calc" value was computed on this page from the rows above it. No value is grade A or B.

## When did Harbor launch, and how did it grow? Timeline

From a company started in mid-2025 to seven trials on its software and a first full-service contract, in about a year. Each entry carries its grade.

- **June 2025** (grade C [9]): Albert Cai and Nathan Leung start the company, then called Harbor Labs.
- **19 September 2025** (grade C [13]): First blog post, on keeping each software release validated for regulators.
- **13 October 2025** (grade C [8]): Magic Capture goes live: AI reads scanned clinic records and fills the study forms, with confidence scores and human review.
- **9 November 2025** (grade C [9]): Partnership with Solaris Research Institute, a pain and spine research site, to run the data capture and monitoring of its upcoming studies.
- **16 March 2026** (grade C [3]): First Y Combinator launch post: an EDC 'live in five clinical studies', launched 'in just eight months'.
- **17 March 2026** (grade C [10]): Case study: a 1,600-couple study goes live nine days after the protocol was handed over, for a flat build fee of US$2,000.
- **29 May 2026** (grade D [1]): Second launch post: seven trials and US$15,600 in monthly software revenue, the founders said, and a first seven-figure contract to run a trial end to end as a CRO.
- **7 August 2026** (grade C [13]): A hands-on Harbor practice lab with the Clinical Research Learning Network for people training in clinical research.
- **November 2026** (grade C [4, 5, 2]): The website still sells the EDC with published prices; Y Combinator describes Harbor as an AI-native CRO with 4 people.

## Why did Harbor work?

### A founder who did the job

The CEO helped run the pivotal trial of a medical device startup and, in his words, lived at a clinical site for a month while staff spent hours entering data by hand. The product attacks those exact steps: building the database, typing records into it, and checking the entries against the paper.

### AI with a person on every value

Magic Capture extracts each value from a scanned page, links it to its place on the page, and gives it a confidence score. A coordinator confirms or fixes the value. The company says every action stays attributable to a human user, which is what an FDA auditor asks for.

### A flat price against long quotes

Large vendors quote weeks of setup work. Harbor publishes a starting price of US$2,000 a month for a commercial study and says it does not charge per subject, per site or per form. A case study names a US$2,000 flat fee for a nine-day build.

### Selling one deal at a time

The customers are sponsors and sites, reached person to person: Harbor wrote that it reached out to NOVA Genomics, the CEO flew to Boston to close the first CRO contract, and both launch posts ask readers for introductions to clinical operations leaders.

## How it was built

These details come from the company's own blog, engineering posts and subprocessor list. They are grade C.

- A full-stack TypeScript codebase: Next.js frontend, Express backend, Passport.js for login, tRPC for routes and Drizzle ORM for the database, as of an engineering post from November 2025.
- Postgres with a constrained entity-attribute-value model, because every trial has a different set of fields, and Zod checks in the application layer (engineering post, March 2026).
- Audit logging written in TypeScript on every write to Postgres, without database triggers, because a missing audit record breaks Part 11 (engineering post, May 2026).
- AI services from Google (Vertex AI/Gemini) and OpenAI, document extraction from Reducto and Extend, Temporal Cloud for workflows, hosting on Google Cloud and AWS, Cloudflare in front, Sentry and Amplitude for monitoring (subprocessor list, effective 25 May 2026).
- A validation package with risk and traceability matrices comes with each commercial license, so a sponsor does not have to write it from zero.

Sources: [14, 13, 7, 4]

## Who tried this before Harbor, and why did they fail?

Four companies in our database of failed startups sold software into health care or tried to be a platform for a whole industry, and failed. The values below are from that database and are not independently verified.

The failed startups come from our database of dead startups, retold here in our own words; Its entries are AI-assisted summaries of public sources, so they carry no evidence grade.

### PatientBank (2016 to 2019; funding US$250K; cause: No Market Need)

- Promise: Patients request, store and share their medical records online.
- Why it died: Complex health care market and strict rules slowed growth; data exchange is driven by institutions, not patients, and Epic and Cerner were built into the hospitals.
- Record: 

### Pear Therapeutics (2013 to 2023; funding US$315M; cause: Unit Economics)

- Promise: FDA-cleared apps that doctors prescribe instead of pills.
- Why it died: Software-style pricing met drug-style costs: FDA clearance did not bring steady insurance payment, and payers had to be won one by one.
- Record: 

### Weidong Cloud (2012 to 2024; funding US$300M; cause: No Market Need)

- Promise: One software platform for Chinese hospitals, doctors and drug companies.
- Why it died: It sold to hospitals, doctors, drug companies and patients at once, never chose one customer, and its product became a mix of conflicting needs.
- Record: 

### Pando (2018 to 2024; funding US$35M; cause: Unit Economics)

- Promise: Real-time supply chain visibility software for shippers.
- Why it died: A general platform for a problem that differs by industry; long implementations, data integration work and a services-heavy model ruined the unit economics.
- Record: 

## What changed

The predecessors failed on two things that Harbor also faces: health care buyers move slowly and follow rules, and software that needs a lot of human work around it stops having software margins. Pando died of a services-heavy model; Harbor has chosen to sell services on purpose, as a CRO, and says its own software will make that work cheaper than at older CROs.

What differs is the narrow job and the regulation that defines it. An EDC has one clear customer, the trial sponsor, and one fixed set of duties under 21 CFR Part 11, so the product does not have to serve every party in health care. And AI that reads messy documents became good enough by 2025 for the company to automate data entry while a person still confirms each value, which the founders present as the reason a four-person company can compete with large vendors.

## Can one person copy Harbor now?

**Verdict: Only with years inside the industry.**

The method needs domain knowledge that one person rarely has: the FDA rules for electronic records, a validated release process, and contacts at sponsors and sites. The founders had a trial operator and an engineer from large software companies. The pattern is still open: find a regulated workflow where people retype documents into a fixed database, and sell AI extraction with human confirmation and the audit trail the regulator requires.

1. Work inside the regulated process first, and write down which steps are manual retyping.
2. Build the audit trail, access control and validation documents before the AI features, because buyers check them first.
3. Keep a person confirming every AI value, and show the source location and a confidence score for each.
4. Publish a flat price against competitors' long quotes, and win the first customers through introductions.

**What is still open.** Other record-heavy regulated work: laboratory records, medical device complaint files, quality management in manufacturing, and the free tier for university studies that now use older free tools.

## What could go wrong for Harbor?

- **What is not verified.** The headline of US$15,600 a month in software revenue rests on one statement by the founders, in their Y Combinator launch post of 29 May 2026. The Y Combinator company page, also written by the company and not dated, gives about US$187,000 in contracted software revenue over twelve months, which is the same figure times 12, so it does not confirm it. No payment-processor profile, filing or press report was found. The US$1.93 million CRO contract is a contract value over three years, stated by the company, not revenue received. A reader could look for a revenue figure in a later interview, a funding announcement or a customer's public filing.
- **Software company or service company.** Since May 2026 Y Combinator's page calls Harbor an 'AI-native CRO', and the first service contract (US$1.93 million over three years, the company said) is larger than the software revenue. Service work needs project managers, monitors and data managers, and it grows with people, not with software. The website still sells the EDC on its own with published prices, and this page covers the software figure only. Later figures may mix the two.
- **Regulation.** Data from FDA-regulated trials must meet 21 CFR Part 11 (electronic records and signatures) and Good Clinical Practice, and patient records fall under HIPAA. Harbor writes that its software 'supports' Part 11 and that 'compliance is a shared responsibility' between the software, the sponsor's procedures and validation; its blog says responsibility 'ultimately rests on the trial sponsor'. An error in AI-extracted data would land in a regulatory submission, which is why each value needs human confirmation. The product makes no treatment or health claims to patients; it is sold to companies and research sites.
- **Competition from established vendors.** Harbor's own comparison pages name Medidata Rave, Veeva EDC, Viedoc, Medrio and REDCap. It describes Medidata Rave as an established enterprise EDC inside a broader Clinical Cloud with separate products for supply, imaging and safety. Large sponsors already have contracts and validated processes with these vendors, which is why Harbor targets small and growing device, biotech and pharma companies.
- **Dependence on model providers.** The subprocessor list names Google (Vertex AI/Gemini) and OpenAI for AI services and Reducto and Extend for document extraction. A change in a provider's model, price or data terms would reach a validated system; the company states zero data retention and business associate agreements with all partners, but no source says how it revalidates when a model changes.
- **Testimonials close to the founders.** One of three home-page testimonials is from a clinical research associate at Biolinq, the CEO's former employer. The other two are from the CEOs of two customers named in a case study.

## Sources

1. Y Combinator Launch, 29 May 2026: Harbor: AI-native CRO for clinical trial operations — https://www.ycombinator.com/launches/Qb6-harbor-ai-native-cro-for-clinical-trial-operations (grade D). Read in full. Founders' post. Source for the seven trials, the US$15,600 in monthly software revenue, the first seven-figure CRO contract and the Boston trip.
2. Y Combinator company page: Harbor, AI-native CRO for clinical trial operations — https://www.ycombinator.com/companies/runharbor (grade D). Read in full on 5 November 2026. Company-written description: contracted software revenue, the US$1.93 million contract, team size 4, founded 2025, Spring 2026 batch, founder names and titles.
3. Y Combinator Launch, 16 March 2026: Harbor: AI system of record for clinical trial data — https://www.ycombinator.com/launches/PiF-harbor-ai-system-of-record-for-clinical-trial-data (grade C). Read in full. Founders' first launch post. Source for the five studies, Magic Build and Magic Capture, and the founders' background.
4. Harbor home page: The intelligent EDC — https://runharbor.com/ (grade C). Read in full on 5 November 2026. Product claims, testimonials, compliance stack and plan summary.
5. Harbor pricing page — https://runharbor.com/pricing (grade C). Read in full on 5 November 2026. Source for the pricing table and the pricing questions and answers.
6. Harbor company page: Meet the agentic EDC — https://runharbor.com/company (grade C). Read in full. Source for the founders' roles, the target customers and the product description.
7. Harbor subprocessor registry, last updated 19 October 2026 — https://runharbor.com/security/subprocessors (grade C). Read in full. Source for the AI, cloud and document extraction providers.
8. Harbor blog, 13 October 2025: Magic Capture: Reducing Clinical Trial Data Entry Burden by 90% — https://runharbor.com/blog/2025-10-13-magic-capture (grade C). Read in full. Source for how Magic Capture works and the company's four-coordinator test.
9. Harbor blog, 9 November 2025: Introducing Our Partnership With Solaris Research Institute — https://runharbor.com/blog/2025-11-09-solaris-research-institute-partnership (grade C). Read in full. Source for the start date ('Five months ago') and the first named site partnership.
10. Harbor blog, 17 March 2026: MotilityCount ApS Launched Their Clinical Study in Just One Week with Harbor — https://runharbor.com/blog/2026-03-17-motilitycount-case-study (grade C). Read in full. Source for the nine-day build, the US$2,000 flat build fee and the outreach to the customer.
11. Harbor blog, 16 April 2026: Why is 21 CFR Part 11 So Important for Clinical Trials? — https://runharbor.com/blog/2026-04-16-21-cfr-part-11-compliance (grade C). Read in full. Source for the company's account of sponsor responsibility under Part 11.
12. Harbor: Medidata vs Harbor comparison page — https://runharbor.com/compare/medidata (grade C). Read the opening and platform-scope sections. The company's own comparison, not neutral.
13. Harbor blog and engineering index — https://runharbor.com/blog (grade C). Read in full. Source for post dates, the engineering posts on Postgres data modeling and audit logging, and the CRLN practice lab.
14. Harbor engineering, 24 November 2025: Why We Love Functional Programming but Don't Use Effect-TS — https://runharbor.com/blog/2025-11-24-why-we-dont-use-effect-ts (grade C). Read the sections on the existing stack. Source for Next.js, Express, Passport.js, tRPC, Drizzle ORM and the TypeScript libraries.
15. Database of failed startups, local export — /method/#dead (grade C). Local export. AI-written summaries, not independently verified. Source for the failed predecessors.

Contact: aiolum@livyden.com. Run Harbor? Send corrections or a verified figure to aiolum@livyden.com (subject: Correction: Harbor).