← Back to Portfolio
AI Development
AI Clinical Assessment Platform
Healthcare client — under NDA12 assessment domains
AI / LLMLocal LLM (Ollama)ReactTypeScriptNode.js / FastifyPostgreSQLKubernetes
An AI assistant that walks a person through a full clinical assessment — built for ADHD and autism screening, and designed so the hard part (getting complete, usable evidence) actually gets done.
A conversational avatar conducts a virtual interview across roughly 12 domains. The important design decision is that the interview is directed: rather than marching through a fixed script, the assistant steers the conversation until it has gathered the evidence a given domain requires, and only then moves on. That's what separates a chatbot from an intake instrument — an incomplete answer doesn't just get recorded, it gets followed up on.
From there the system produces AI-generated report overviews for the reviewing clinician and applies pattern recognition to raise red flags worth a closer look. To be exact about what that means: the platform surfaces suggestions for the clinician — never diagnoses. It assists the doctor's judgment; it does not replace it, and it was built that way on purpose.
Delivered end to end as a configurable multi-clinic product, so each practice can tailor its own flows and branding: React and TypeScript on the front, Node and Fastify APIs, a PostgreSQL data model for assessments and sessions, and a local LLM running the interview engine and report generation — which keeps sensitive clinical conversation off third-party infrastructure. Deployed on Docker, Kubernetes, and AWS.
A conversational avatar conducts a virtual interview across roughly 12 domains. The important design decision is that the interview is directed: rather than marching through a fixed script, the assistant steers the conversation until it has gathered the evidence a given domain requires, and only then moves on. That's what separates a chatbot from an intake instrument — an incomplete answer doesn't just get recorded, it gets followed up on.
From there the system produces AI-generated report overviews for the reviewing clinician and applies pattern recognition to raise red flags worth a closer look. To be exact about what that means: the platform surfaces suggestions for the clinician — never diagnoses. It assists the doctor's judgment; it does not replace it, and it was built that way on purpose.
Delivered end to end as a configurable multi-clinic product, so each practice can tailor its own flows and branding: React and TypeScript on the front, Node and Fastify APIs, a PostgreSQL data model for assessments and sessions, and a local LLM running the interview engine and report generation — which keeps sensitive clinical conversation off third-party infrastructure. Deployed on Docker, Kubernetes, and AWS.
Screenshots
