Aegis AI Clinical Decision Support — Multi-Model LLM Orchestration Platform

Giannis Roussos
Published

0
  • Project creator

Multi-model AI clinical decision support platform built independently in both Next.js and Ruby on Rails to benchmark architectures.

What it does
Clinicians paste medical records and receive distilled summaries, suggested differential diagnoses, treatment options, and live evidence from PubMed — all grounded in the document context only, with no external knowledge injection (hallucination eliminated by design).

Architecture highlights

  • Five-model Gemini/Gemma waterfall cascade (Gemini 2.5 Flash → Flash Lite → three Gemma variants)
  • Reactive 429/503 switching with capability detection between Gemini and Gemma model families (Gemma does not support the systemInstruction field — handled with a runtime supportsSystemInstruction() check)
  • 45-second AbortSignal timeout per call
  • RAG via Supabase pgvector with a pure-TypeScript cosine-similarity fallback path for RPC failures
  • HL7 FHIR R4 data explorer and raw .hl7 file ingestion alongside PDF and text
  • Live PubMed evidence grounding for clinical claims
  • Redis rate limiting, Zod schema validation, structured audit logging

The Rails benchmark
Rebuilt the entire system in Ruby on Rails over a single weekend (framework learned from scratch) to compare architectures. Result: 60% faster total execution (5.6s vs 14.7s), 85% reduction in JavaScript payload.

Stack: TypeScript, Next.js (App Router), Ruby on Rails 7, Supabase (Postgres + pgvector), Vercel AI SDK, Redis, Zod, Hotwire (Turbo/Stimulus), Vercel, Fly.io

Live demos:

Source: https://www.twine.net/signin