Sheikh Babar A.

Most projects that use a model fail the same way. The demo works on five clean invoices, then production sends a scanned fax, a merged-cell spreadsheet, and a form nobody has seen before, and suddenly there is no way to tell which extracted fields are actually trustworthy. I build the version that survives that. Per-field confidence scoring, a human review queue for anything below threshold, and a schema per document type so the output is predictable enough to push straight into your system. What I work with: - Document and data extraction: OCR, PDF, DOCX, XLSX parsing, structured JSON output with typed schemas - LLM pipelines: OpenAI, Claude, RAG, vector search, agent workflows, tracing and evaluation - Backends: TypeScript, Node, Bun, Python, PostgreSQL, background job queues - Frontends: React, Next.js, TypeScript - Scraping and automation: Playwright, Puppeteer, scheduled crawlers I write tests against real infrastructure rather than mocks, and I will tell you when a project should be scoped smaller before it should be started. Send me what you are trying to extract and what you need to do with it downstream, and I will tell you whether it is a clean fit.