Ethan Haas

I build document-extraction pipelines, AI systems, and full-stack SaaS for finance and regulated-document work, and I make the output verifiable rather than just plausible. Tested code does the math, not the model.

Over the past year I shipped five production document-automation tools for a public accounting firm, replacing manual, error-prone audit workflows. Deterministic PDF-extraction pipelines turn unstructured financial reports into structured Excel, validated by automated tie-out checks and golden-file tests, and delivered as single-file Windows apps and FastAPI/React web apps for non-technical staff. The most involved is an 8-stage LangGraph agent pipeline on Gemini Vision that reads long unstructured documents and extracts, classifies, validates, and dedupes every actionable item into a workbook.

My LLM work follows a model-proposes, tested-code-decides approach: structured output, source citation, and human-in-the-loop review so low-confidence results go to a person rather than quietly shipping. I also made a real local-versus-frontier routing decision under a confidentiality duty, keeping four of those five tools fully offline while exactly one calls a frontier model, by explicit client opt-in.

Alongside that I build and operate a live B2B tax-compliance SaaS, multi-tenant with billing. Earlier I built a RAG document-retrieval system that lets attorneys query legal text with precise page-level citations, and led digital strategy for a legal-services venture that reached profitability in its first months.

The clearest example of how I think is a benchmark I published where I trained and compressed neural networks, built an exact solver for ground truth, and then reported that a 0-byte handcrafted search beat every model I trained. I would rather ship the true answer than the flattering one.