Building AI systems for healthcare operations.
Two years embedded with Medicaid providers as one of two engineers on a platform that reached 35,000+ reported enrollments and re-enrollments. Agentic systems for prior authorization and clinical data, shipped with eval harnesses and honest numbers.
19.2%
cohort engagement
~12.8x the published Medicaid SMS benchmark
93.7%
macro-F1
on a held-out set of synthetic cases
I publish the failures
honest evals
The LLM judge that failed its own agreement checks and the fine-tune that lost out-of-distribution are both in the repos.
Featured projects
Systems I built and evaluated
Healthcare-AI experience
Two years at Trust.care
At Trust.care, I was one of two engineers building an AI-powered referral platform embedded directly with Medicaid providers and community-based organizations.
The population set the design: 63.8% of the cohort had no email, so the system was SMS-first by necessity, not by preference.
It reached a ~955K-member Medicaid cohort across 100+ CBOs and processed 5.12M+ outreach messages, with a reply-classification pipeline separating high-signal responses from noise.
The work contributed to 35,000+ reported Medicaid enrollments and re-enrollments, at 19.2% cohort engagement, roughly 12.8x the published Medicaid SMS benchmark.
5.12M+
outreach messages
to a ~955K-member Medicaid cohort
118K+
replies classified
by intent, high-signal separated from noise
Engineering philosophy
How I work
Build next to the user.
Forward deployed by instinct. The clearest requirements come from sitting with the people who live the problem, not from a spec.
Evaluate honestly, publish the failures.
Every project ships with an eval harness and the numbers that did not work, including the ones that embarrass the demo.
Guardrails before autonomy.
Human approval gates and deterministic checks come first. An agent earns the right to act by proving it can be stopped.
One size fits none.
Design for the actual population. A cohort where most members have no email does not want to be reached by email.