Quan D.

I build production AI pipelines that turn messy documents into structured data. 8 years in Python and data engineering. I specialize in document processing systems — the kind where you throw in a stack of scanned PDFs and get clean, validated JSON out the other side. ## What I actually build ### LLM-Powered Extraction Pipelines Multi-agent systems using Claude, GPT-4, and GPT-4o with structured output enforcement (Instructor, Pydantic). I design domain-specific agents that classify documents, extract fields, validate data, and reconcile conflicts across sources. Not prompt-and-pray — schema-enforced, retry-aware, production-grade. ### OCR & PDF Processing Amazon Textract (async API), PyMuPDF, PDFPlumber, Camelot, Azure Document Intelligence. I handle the real-world mess — scanned documents with poor quality, mixed orientations, concatenated multi-document PDFs that need splitting and boundary detection. Table extraction, form parsing, layout-aware chunking. ### Serverless & Event-Driven Architecture AWS Lambda, SNS, SQS, Step Functions, EventBridge, ECS/Fargate. I build pipelines where S3 uploads trigger processing chains — fan-out to parallel extractors, dead-letter queues for failures, CloudWatch monitoring for quality metrics. ### Healthcare Document Systems Currently running a production pipeline that processes medical records daily — licenses, clinical notes, immunization records, provider credentialing documents. Full pipeline: PDF ingestion → OCR → document classification → LLM extraction → deduplication → normalization → structured output. Built with HIPAA considerations. ### Legal Document Systems Production pipeline for UK cost-assessment workflows — ingests multi-document bill bundles (PDFs, scans, Excel ledgers, Outlook .msg), classifies them, extracts disbursements, receipts, invoices, and party data, reconciles figures across sources with a 5-case matching algorithm, and generates court-compliant inter-partes narratives. Every extracted figure traces back to its source page for audit defensibility. ### Speech-to-Text Pipeline - AWS Transcribe Medical (streaming WebSocket + batch S3 jobs) with HIPAA-compliant configuration - Deepgram integration as an alternate ASR backend - Speaker diarization (up to 30 speakers), stereo channel separation, custom medical vocabularies, specialty-specific vocab lists (e.g. orthopedics) - Multi-threshold confidence analysis for dot-phrase and quick-text detection ### LLM Clinical Note Generation - Instructor + Pydantic schema-enforced structured outputs (Claude / GPT-4 / Gemini) - Prompt Studio: a template-authoring system where clinicians design reusable note templates with variables, segment bindings, and inline citations back to the raw transcript - Two output modes — Ambient (multi-category SOAP-style synthesis) and Dictation (verbatim with cleanup) - DOCX template engine with inline placeholder tokens, dual-archive export, and anchor-based fill for exact-fidelity Word output ### AI Agent Orchestration LangChain, LangGraph, MCP servers, RAG systems, function calling. I build agents that do real work — not chatbots that summarize, but extraction systems that produce validated, structured data from unstructured sources. ## Anthropic API Cost Optimization I help teams cut Claude/LLM API spend in half on production deployments without sacrificing product quality. I work across the levers that actually move the bill: - **Prompt caching** — designed around real traffic patterns, not just toggled on - **Model tiering** — Sonnet where reasoning matters, Haiku everywhere else - **Context discipline** — lazy-loading tools, skills, and instructions instead of dumping everything upfront - **Tool-call efficiency** — fewer round-trips, smaller payloads I've done this on a medical-document extraction pipeline and a LangGraph trading agent — both running in production with real users and real budgets. If you've inherited an over-budget Claude deployment, I know where to look first. ## Tech I use daily Python, FastAPI, Anthropic/OpenAI/Gemini, Instructor, LangChain, LangGraph, Textract, PyMuPDF, PostgreSQL, Redis, S3, Lambda, SNS/SQS, Docker, Pydantic. 20+ projects delivered on Upwork. I'm strongest when the problem involves turning unstructured documents into clean, structured data at scale.