Prototype
GenAI Medical Assistant — RAG Platform (Safety + OCR Intake)
Reliable AI assistant chatbot: grounded answers, eval scripts, and approval workflow
Role
Solo Developer
Type
AI Platform (Web + Backend)
Stack
Ollama (Local LLM)ChromaDB (Vector DB)Docker ComposeGitHub ActionsOCR PipelineFastAPIPythonRAG (Retrieval-Augmented Generation)Next.js
Overview
A production-style GenAI assistant that answers using retrieval-grounded context rather than guessing. I built a full pipeline: document intake (OCR + scan quality scoring), review/approval workflow, RAG retrieval, and a deterministic safety/triage layer. The system is containerized with Docker Compose and backed by CI (lint/test/build) to keep it reproducible and deployable.
Key Features
- Retrieval-grounded answers (RAG) to reduce hallucinations
- Deterministic safety + triage layer before responding
- OCR document intake with scan-quality scoring
- Review/approve workflow for document sources
- Evaluation scripts (grounded vs non-grounded comparisons)
- Docker Compose orchestration with health checks
- CI pipeline (lint/test/build) for reproducible deployments
Challenges & Learnings
- Designing safety rules that are predictable (not model-mood based)
- Making the stack reproducible across machines with containers
My Contributions
- • Designed end-to-end architecture (frontend, API, retrieval, storage)
- • Implemented OCR ingestion + scoring + approval workflow
- • Built retrieval pipeline and response assembly logic
- • Added deterministic safety/triage checks
- • Containerized services with Docker Compose + health checks
- • Added GitHub Actions CI for lint/test/build gates
- • Wrote evaluation scripts to compare grounded vs non-grounded outputs