UrbanCare AI: AI-Native Clinical Ecosystem for Mechanistic Interpretability

The problem

Most clinical AI is a black box — physicians can't audit how a diagnosis was reached, so they can't trust it and compliance teams can't sign off on it.

My approach

A Semantics-to-Vector interface that bridges human-readable clinical semantics and internal model vectors, making every layer of the 30-agent pipeline inspectable — from lab-report extraction to differential diagnosis.

What I built

A 30-agent, 8-layer clinical pipeline built on MedGemma + TxGemma + Gemini. Provides a Semantics-to-Vector interface so physicians can inspect how medical AI models reach their diagnoses — from lab report extraction and X-ray fracture detection to differential diagnosis and voice copilot.

How it works

Production-ready AI-native clinical dashboard acting as a Doctor Consultant. Leverages Google HAI-DEF open-weight models (MedGemma 4B for vision/document tasks, TxGemma for therapeutic reasoning). Features a transparent Semantics-to-Vector interface bridging human-readable clinical semantics and internal model vectors — making AI reasoning auditable at every layer. Multi-tenant hospital platform with role-based access control, real-time vitals SSE streaming, and voice clinical copilot powered by RAG over patient memory vectors.

What I'd do differently

Interpretability isn't a feature you bolt on at the end — for clinical tools, it has to be the interface itself.

Good questions to ask

3 questions an interviewer could ask about UrbanCare AI

Why did you treat interpretability as the interface itself rather than a feature bolted on at the end?
Based on: Interpretability isn't a feature you bolt on at the end — for clinical tools, it has to be the interface itself.
How did you bridge human-readable clinical semantics and internal model vectors to inspect the pipeline?
Based on: A Semantics-to-Vector interface that bridges human-readable clinical semantics and internal model vectors, making every layer of the 30-agent pipeline inspectable — from lab-report extraction to differential diagnosis.
How did you address the lack of auditability that prevents compliance teams from signing off on diagnoses?
Based on: Most clinical AI is a black box — physicians can't audit how a diagnosis was reached, so they can't trust it and compliance teams can't sign off on it.

AI-written from this site's content

AI-generated · may be wrongHow it works Privacy

At a glance

Technologies: 8
Language: TypeScript
Updated: Feb 2026

React 19
Node.js
PostgreSQL
MedGemma
TxGemma
Gemini
LangChain
GCP

#medical-ai
#interpretability
#multi-agent
#healthcare

Compare with another project

Compare with

Related work

AI-generated · may be wrongRanked by meaning with Gemini embeddings of each write-up. Semantic is cosine similarity; 1 would be identical.How it works Privacy

UrbanCare AI