built AI products in life sciences for 12 years. now building my own. dad. SF.
San Francisco, CA
Joined November 2014
- Tweets559
- Following1.7K
- Followers314
- Likes447
Trying something new: short videos on X.
I've been overthinking making content, so I gave myself 60 minutes to shoot, edit and post. The goal: get the juices flowing
Right now I'm just excited about flying cars. Wanted to capture that feeling.
Got accepted into Claude for Startups. Then got unaccepted. Fastest startup exit I’ve had.
@AnthropicAI
9/ Next: prove it transfers to real homes. Occupational therapists will label real, consented, de-identified homes, and we'll check whether the models that fail in the sim also fail in the real world.
Building AI for the home? I'd love to talk.
10/ Try it: healthdojo.dev
Thank you @pearvc, @NEA, @CathayInnov, @OpenAI, @awscloud and @jpmorgan for putting this on, and @elijahyi, @shrutimerch and the judges (@dr_alphalyrae )
Great day!
8/ Built solo in one day, by directing a team of AI coding agents working in parallel: pipeline, 3D worlds, agent environment and site.
A year ago this would have been a team and months.
7/ Every labelled room also becomes a walkable 3D world. Drop a model in, give it 8 moves, see if it finds the hazard.
GPT-5.6 kept a loose bath mat in view for six steps, flagged five other things, and never flagged the mat. 👇
2/ AI is moving into patients' homes: robots, smart glasses, care apps that scan your house from a phone.
There's no safe way to test it. Real homes have no labels, and the rare dangerous cases almost never show up. So teams find the failures in pilots, on real people.
5/ Then it gets harder. Five levels: one obvious hazard, then subtle ones (a towel bar used as a grab bar), then safe look-alikes that should NOT be flagged, then clutter and dim light.
In a first run through all five levels, every model scored lower at level 5 than at level 1.
4/ How it works:
• Start from a real clinical guideline (CDC STEADI)
• Every line becomes a rubric row, with the source quoted
• Generate a clean room, then paint the hazard in
Each hazard is labelled before the image exists, then checked by a vision model.
3/ Why I built it: I've been building multimodal AI for home health, and I had no way to pre-validate my models before real-world testing.
No labelled home data. No clinical test environment. Nothing between "works on my laptop" and "works in a patient's home."
So I built one.
1/ Made the finals at Saturday's Healthcare AI Hackathon, hosted by @pearvc, @NEA and @CathayInnov.
I went solo and built HealthDojo: a flight simulator for home-health AI.
Here's what it does, and what 23 AI models got wrong 🧵
This makes me nervous in the Codex app.
I’d love a guardrail that says: “I’m about to start a big job. Want to move this to a new thread first?”
@jxnlco @thsottiaux @OpenAIDevs
On days like these, I’m thankful for capitalism and competition.
I can actually switch instead of getting locked in. Short-term pain - Opus is that damn good - but OpenAI will catch up, prices keep dropping, and at the end of the day we the consumers win.
The market decides. Fast.
@steipete
woke up and my mentions are full of these
Both me and @davemorin tried to talk sense into Anthropic, best we managed was delaying this for a week.
Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source.
I genuinely want to start learning how to contribute to open source projects now- just an incredible outcome for everyone- congrats @steipete!!!
Peter Steinberger is joining OpenAI to drive the next generation of personal agents. He is a genius with a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people. We expect this will quickly become core to our product offerings.
OpenClaw will live in a foundation as an open source project that OpenAI will continue to support. The future is going to be extremely multi-agent and it's important to us to support open source as part of that.
DOGE just open-sourced the largest Medicaid dataset in HHS history.
227M rows. $1.09T in payments. 2018–2024.
I ran 5 fraud-pattern SQL checks on it in ~20 minutes using @openclaw + DuckDB.
Here’s what popped. 🧵
6/
⚠️ These are anomaly signals, not accusations.
Some may have legitimate explanations.
Point is: this is basic SQL (GROUP BY, ratios, YoY checks).
Once data is public, these patterns are fast to surface.
7/
Queries (copy/paste):
gist.github.com/pistachiopra…
Dataset:
opendata.hhs.gov/datasets/me…
Ran with OpenClaw + DuckDB on the 2.9GB parquet in ~20 minutes from my phone.
