An Update of the Past Four Months.

The Flour Mill

The three biggest things I've been working on: PMF through potential customer interviews, building the agentic AI infrastructure, and getting each agent's specific functionality reliable.

For PMF this has mostly happened through cold calls, cold emails, and just being a yapper. The yapping is fun. It needs to be focused, but I've met and connected with so many cool people. Even if this thing doesn't work out, I built a great network and learned about a lot of cool problems. They were hyperspecific, but I think that's the nature of trying to sell to small businesses (that's a whole other conversation).

For the agentic infrastructure, this has posed some issues. There are a variety of electronic health records and the model needs to be ready to handle any of them. Some knowledge can be shared between them but overall. It's very individualistic. These are real moats that I can cross with time but oh boy!

The biggest GTM for MAIA would be an efficient medical coder agent. A lot of these physicians outsource their coding and billing services. I've even been told some physicians make their own companies overseas in order to have cheaper medical coding done.

Pulling this off has overall been difficult. There has been some progress, but not as much as I had hoped.

I measure my results on a public test set that academic research uses for this exact problem. The headline number is "first try accuracy": when the system suggests one best code for a case, how often is it exactly right? We are now at about 46 out of every 100. We started at 36, which is roughly where the major general-purpose AI assistants land when you ask them to do this cold. Across six iteration cycles we have moved that 36 up by about ten absolute points, which is a 27 percent relative gain.

For context: the biggest, most quoted general purpose AI models score somewhere between the high 20s and mid 40s on this same task. The published academic baseline that researchers in this space cite sits at about 40. We are now slightly above it. Research's current state of the art on the same dataset is around 60, but it leans on a couple of techniques we have not implemented yet, and that gap is well-understood.

One specialty in particular, cardiology, lands at about 80 first try accuracy.

Fires Always Stop Burning

I'm not defeated by this.

The plateau is exactly where small teams get stuck on this kind of problem, and the path off it is pretty well known in other domains. The wrinkle here is that an agent failure in healthcare lands a lot harder on the end user than it does in most software, which makes the usual "ship the worst version and iterate" advice harder to follow. The underlying recipe hasn't really changed though. I still need this in front of real users, even if the pilot has to be tightly chaperoned.

Second, the system is already wired for the kind of "continual" learning this needs. It produces flagged suggestions with confidence scores and surfaces what it isn't sure about. Bolting on a correction loop is all about signing with a willing partner.

Third, cardiology specific notes are hitting ~80%. This tells me the architecture is in the right neighborhood. If I narrow the scope to a specialty where I already have enough of the right kind of data, the agent already works at a usable level. This could be my starting point for the next partnership.