Blood Report Analyzer (AI Health Agent)
Subconscious AI x ACM Hackathon, Finalist · hackathon project
A hackathon prototype that uses an agent and typed tools to analyze blood-report fixtures. I built backend and agent integrations.
IllustrationHighlights
- Three JSON-Schema-typed tool endpoints on a Hono server that the Subconscious AI (TIM-GPT) agent calls, choosing tool order rather than following a fixed script
- 47-biomarker reference database across 8 clinical categories and a 28-entry recommendation knowledge base
- Five-level severity classification that checks critical thresholds before low/high bands
- Verified on the committed fixture: 26 of 26 biomarkers extracted, 21 urgency-ranked recommendations generated
- Next.js UI, which runs on mock data rather than the live pipeline
How it was built
For the Subconscious AI x ACM Hackathon, our team built an AI blood-test analyzer. I owned the backend and agent: a Hono server in TypeScript exposing three tools, one to extract biomarkers, one to classify them against a 47-biomarker reference database, and one to generate recommendations. The Subconscious TIM-GPT agent decides which tool to call and in what order, with a direct-call fallback when no key is present. On the committed sample report, 26 of 26 biomarkers were extracted and 21 recommendations produced. Limits: the Next.js interface runs on mock data, and nothing was fine-tuned.