Aaron Li

About

LLM/document AI engineer who owned a production-grade contract extraction pipeline at CORAMA.AI, ingesting PDFs and dynamic JavaScript sites from 1,000+ government sources. Built a hybrid deterministic+LLM system with two-phase prompting, Pydantic guardrails, confidence scoring, and human-in-the-loop review—cutting error rates from ~35% to <5% and processing 50k+ documents at ~95% accuracy. Also built clinician-in-the-loop orchestration in research, reducing manual labeling time from 3–4 hours to ~50 minutes.

Key Strengths

• Built and owned end-to-end production document extraction pipeline across 1,000+ heterogeneous government sources
• Designed hybrid extraction system (deterministic regex + LLM for unstructured text) with schema validation guardrails
• Reduced LLM error rate from ~35% to <5% by decomposing tasks into two-phase prompting and adding validation + human review
• Scaled processing to 50,000+ documents with ~95% accuracy using confidence scoring and human-in-the-loop for edge cases
• Strong orchestration judgment: uses LangChain for standard workflows and custom asyncio for high-control, high-throughput batch jobs
• Evaluation-driven development: maintained ~1,000-doc labeled test set and ran automated evals on prompt changes to prevent regressions
• Effective cross-functional collaboration translating vague stakeholder needs into concrete entities/requirements (business ops, customers, clinicians)

Experience

Education

Skills

Languages

Publications

3 publications