We lost on vendor maturity

One of the executives told us he loved it. Then he asked how he was supposed to trust two students with a multimillion-dollar contract.

That was the meeting. Everything before it was preamble, and everything after was arithmetic.

The problem was boring, which is why it was worth solving

The orthopedics department at Yale Health receives thousands of faxed documents a month. Referrals, authorizations, forms that need a physician's signature. Two full-time staff existed to read the incoming fax, work out which patient it belonged to, work out which doctor was treating that patient, and forward it to the right administrator. That was not part of their job. That was their job.

The administrator opened the document in Adobe and manually added text boxes and checkboxes to a scanned page. They looked the patient up in Epic. They found the information the form wanted, which was scattered across the chart. Then they sent it through DocuSign to the physician, who signed it or bounced it back. Then it went back to the administrator, back to the routers, and back out by fax to whoever had sent it. Average time from arrival to return: one to two weeks.

Nobody in that chain was doing anything unreasonable. Every step existed for a reason. And the whole process was invisible, because it never appeared as a line item anywhere. It was distributed across people who were nominally doing something else.

That is where patients get lost. A referral sitting in a queue for eleven days is a person who has not been seen.

We were not obviously the right people

I had not written code in years. My cofounder could not have told you the difference between Python and SQL. We met at Yale, became friends, and found this through the university's network.

I want to be honest about how the build actually went, because the version where two students calmly execute a technical plan is not what happened. I started by asking ChatGPT, Gemini and Claude how to do things, on their websites, one question at a time. That turned into Cursor. That turned into Copilot. I learned the shape of the problem by failing at it repeatedly against a corpus of real faxes I put together myself.

The extraction approach is the one part I cannot get into, because it is covered by the NDA with Yale. What I can tell you is what it had to do. Read a fax that has been through a scanner three times. Pull the fields a reviewer actually needs. And be honest enough about its own confidence that a human can catch it being wrong.

That last requirement turned out to be the whole product. In a clinical setting, a value that is wrong but looks certain is more dangerous than no value at all, because it gets trusted. So nothing was ever silently filled in. Everything stayed editable, and the places the system was unsure were the places it asked.

The blocker we could not solve, and the pilot we ran anyway

We could not get access to Yale's Epic instance. Not without SOC 2 and HIPAA compliance already in place, which two students did not have and could not manufacture on that timeline. So we had a working system and no way to prove it worked.

We went to the orthopedics administrators and designed a pilot around the constraint instead of waiting for it to lift. Templates for the documents they saw most often, common patterns pre-filled, patient data still entered by hand, everything editable. For the doctors the win was different: most of what a physician signs is simple and repetitive, so we grouped similar documents into batches with a summary of what they were approving.

Over two months and just over 2,000 documents, we saved the team roughly 300 hours. That works out to about nine minutes of staff time per document. Nine minutes is the number I would defend in any room, because we measured it. We also modeled the annual opportunity across the health system at north of $15 million. That one is a model, it turns on how many departments look like orthopedics, and I would walk you through the sensitivity before asking you to believe it.

The meeting

We cleared security review. We cleared clinical workflow review. We reached the final stage, which was a meeting with the health system's senior digital and clinical information executives. We brought the product, the pilot data, and one ask: access to Epic, and formal approval to test across more departments. The orthopedics team came with us. They validated the results.

The executives agreed. They liked what we had built. And then the conversation turned to us.

Two students wanted a contract with Yale Health. How could they trust us to run a multimillion-dollar engagement? What would stop us from walking away halfway through? I argued that the value was large enough that I would stay with it for years. That was true, and it was not the point. The follow-up was the one I could not answer. If protected health information leaked, what in my history suggested I could handle it? I had never run a company. I had never managed anything at that scale.

They needed an established vendor. Someone with history, and other clients who had survived them. The vendors they preferred did not cover the whole stack we had built.

What I actually took from it

For a while I told this story as though something unfair had happened. It had not.

A health system that requires a vendor who will still exist in five years is not being bureaucratic. It is being correct. The cost of a vendor disappearing mid-contract is not the software. It is the workflow that has already been rebuilt around it, the staff retrained, the integration half-finished, and the patients moving through all of it while somebody works out what to do next. Weighed against that, “these two are talented” is not a reason. It is a hope.

By that point we were not being evaluated on the product. We were being evaluated on whether we would still be there.

Two things I would do differently. Test institutional credibility in week two, not month fourteen: we validated the workflow and the economics, but never whether an institution like this could buy from a company like ours at all. That was answerable early and cheaply, by asking one person in procurement a direct question. We asked it at the end, after we had spent everything getting there.

And get a second design partner with a lower maturity bar. Everything was concentrated in one relationship at one academic medical center. A community health system or a specialty practice group would have moved faster and demanded less institutional history. More importantly, they would have given us the institutional history the first buyer wanted.