Data Scientist
Innovaccer · Oct 2024 – Present
- Architected a 15-node LangGraph multi-agent orchestration system with model-specific routing logic: Claude Sonnet for clinical reasoning, Claude Opus for medical necessity, Gemini 2.5 Flash for high-throughput parsing via TrueFoundry LLM gateway; achieved 94% ICD-10-CM, CPT, and HCPCS extraction precision benchmarked against 500K+ historical enterprise patient encounters.
- Defined and engineered an 8-layer hallucination prevention framework: Pydantic structured outputs with mandatory source-citation fields, cross-agent verification passes, rule-based ICD/CPT validators against 200K+ CMS NCCI rules (MongoDB), LangGraph ReAct self-correction loops, and per-agent LLM-judge evaluation tracking precision, recall, F1, and automated hallucination scoring.
- Designed per-model circuit breakers (5-failure threshold, 60s recovery window, Redis-backed) with exponential backoff and token budget pre-flight guards, cutting failure-cascade recovery time from 210s to under 1s during LLM gateway outages; parallelized independent agents via Temporal workflow orchestration, reducing end-to-end latency 60% (45s to 18s) on 3-page clinical documents.
- Established production observability across all 15 agents via MLflow tracing: logging token consumption, per-agent latency, and prompt aliases across production and staging environments via Databricks Unity Catalog, enabling precise per-client cost attribution and model resolution auditing.
- Architected and owned Incurate end-to-end a production clinical NLP platform for unstructured medical documents: OCR → section-aware chunking → LLM entity extraction → semantic code mapping (ICD-10, CPT, RxNorm) → automated evaluation.
- Fine-tuned MedGemma-27B and Qwen-2.5-7B via LoRA/PEFT on domain-specific clinical datasets, boosting entity extraction F1 from 65% to 94%; selected as the production model after benchmarking against 15+ alternatives.
- Drove the full-stack migration of the clinical NLP pipeline (OCR + extraction + code mapping) from AWS Comprehend Medical to open-source infrastructure, cutting annual platform costs 94% ($90K to $5K/year).
- Designed and optimized a Milvus vector search index over 100K+ medical codes (ICD-10, CPT, RxNorm) achieving 91% mapping accuracy, 3x faster than keyword search; PyMuPDF + Vision LLM OCR pipeline achieving 99% accuracy across multi-page clinical documents with section-aware chunking.
- Architected a microservices backend (FastAPI, AsyncIO, Redis, Celery) reducing inference latency 5.5× for enterprise deployments; system now runs across 3 enterprise healthcare products in production.
- Benchmarked 15+ LLMs (GPT-40, Gemini 1.5 Pro, LLaMA-3, Qwen-2.5) on entity F1, hallucination rate, and cost-per-1K-tokens; model selection framework adopted as the evaluation standard across the enterprise product stack.

