Skip to content
All work

AI systems, 2026

PharmaTrace

Six LLM agents check a medicine against three drug registries and refuse to invent medical terms.

Role
Personal project
Type
AI systems
Links
Source on GitHub
LangGraph agents
6
drug registries
3
voice languages
4

How it fits together

  1. React 19 PWATesseract OCR, Whisper voice in 4 languages, IndexedDB outbox
  2. scan or speak
  3. FastAPIStrict Pydantic schemas and the jargon firewall
  4. 6-agent graph
  5. LangGraphBarcode, FDA lookup, recall, interaction, safety and report agents
  6. asyncio.gather
  7. RegistriesOpenFDA, RxNav / RxNorm, CDSCO, FAERS signals
  8. validated output
  9. Report + auditSeverity-sorted warnings, SHA-256 hash-chained log in Supabase

The problem

Checking whether a medicine is genuine, recalled or unsafe for a particular patient means cross-referencing several registries that disagree in format. An LLM can summarise that well, but it also likes to make up confident clinical terms, which is the one thing a health tool cannot do.

How I approached it

  1. 1

    Split the job into agents

    A LangGraph pipeline with six agents (barcode, FDA lookup, recall, interaction, safety, report) over OpenFDA, RxNav and India's CDSCO registry. Independent lookups run in parallel with asyncio.gather to cut end-to-end latency.

  2. 2

    Replace counts with signal

    Naive adverse-event report counts were replaced with NLM RxNorm clinical mapping, FAERS signal strength and severity-sorted warnings.

  3. 3

    Constrain the model

    Strict Pydantic schemas plus a field-validator 'jargon firewall' structurally block hallucinated medical terminology. Patient age, weight and renal function are injected into prompts to trigger dosage warnings.

  4. 4

    Work offline

    A React 19 PWA with in-browser Tesseract OCR for expiry dates, IndexedDB caching and a Background Sync outbox for poor connectivity.

What I built

  • Whisper-based voice input in Malayalam, Hindi, Tamil and English.
  • Every verification is written to a SHA-256 hash-chained audit log, so tampering with history is detectable.
  • Deployment configs for Railway (API) and Vercel (PWA).

The result

A verification flow that answers in the user's language, keeps working without a network, and returns warnings that are traceable to a registry rather than to the model's imagination.

Built with

  • Python
  • FastAPI
  • LangGraph
  • Pydantic
  • Llama 3.3
  • Whisper
  • React
  • PWA
  • Supabase
  • PostGIS
  • Tesseract