Neurithum AI — AgeCare Platform
CurrentAI Engineer
Python · FastAPI · LangGraph · pgvector · PostgreSQL · Redis · AWS · Docker
Agentic EHR Chatbot (25-tool LangGraph pipeline)
Shipped a production dual-agent system for natural-language querying over Electronic Health Record data, where an LLM intent router (gpt-4o-mini, temp=0) dispatches to a data agent (25 async SQL tools across 11 EHR domains) or a clinical reasoning agent (4 extraction tools). Added real-time SSE streaming, PostgreSQL-persisted memory with sliding-window LLM summarisation, anti-hallucination guardrails (forced first tool call + output blocklist), and PHI audit logging on every data access.
Hybrid RAG retrieval system
Designed a two-stage BM25 + pgvector semantic search with Reciprocal Rank Fusion for clinical entity extraction from free-text notes, achieving a 74% reduction in LLM token cost through targeted retrieval. Built the ingestion pipeline for guideline PDFs using semantic chunking and OpenAI text-embedding-3-small embeddings.
Behaviour pattern analysis — 5-layer ML pipeline
Engineered SQL signal extraction → SciPy KDE clustering to model behaviour frequency distributions → LangGraph reasoning agent for causal hypothesis generation → RAG intervention retrieval → human-review draft.
Clinical deterioration prediction (CDTR)
Computed patient baselines from longitudinal EHR data and detected statistical drift in vital-sign and behavioural signals, auto-generating structured clinical notes and facility-wide dashboards for proactive escalation.
Data pipeline and SQL optimisation
Replaced N+1 Python loops with PostgreSQL set-based views, benchmarked across 18 runs: 2.64× peak speedup, 44.4% payload reduction, and 75% LLM cost savings. Designed 5 AI-specific SQL views powering unified patient timeline, clinical profile, behaviour intelligence, and compliance features.
Care plan intelligence (ICP)
Built a deterministic gap-detection engine, preference scanner, and LangGraph reasoning agent for care plan generation with compliance checking, 4-layer Redis caching, multi-tenant scoping, and RBAC across 5 roles.
Production security and compliance
Implemented JWT/RBAC middleware with PHI access audit logging, OWASP security headers, sliding-window rate limiting, parameterised SQL, and LLM output guardrails blocking unsafe clinical recommendations.