Subhajit B.

I help businesses get back the hours their team spends reading, understanding and extracting data from documents, with every value verified before anyone uses it. One pipeline I built gives a UK consultancy back 150+ staff hours every month, almost a full-time person, and it has run in production for over 2 years. ──────────────────────────── BEFORE AND AFTER Before: someone opens each PDF, finds the figures and types them into a spreadsheet. Mistakes surface weeks later, in a client report. After: a batch of lab reports from 10 labs runs in 10–12 minutes, every value is checked, and the results are in the team's inbox before they log on. ──────────────────────────── WHAT YOU GET • Data you can trust: every field checked against a schema, and anything uncertain goes to a person first • A system that survives month six: each new layout becomes a regression test, so a fix never breaks an old document • No extra steps for your team: emailed documents are picked up every morning, and results land in Excel, an API or your own tools • No surprises: a short written update after every working day on your project, and a weekly Loom walking you through progress ──────────────────────────── RESULTS • UK water consultancy, 2+ years in production: about 5,000 lab-report pages a month, 10 labs, 22 parsers, 150+ staff hours of re-typing removed every month • Same client: tested OpenAI extraction on their real documents for new layouts before anything went live • PDF tables turned into formatted Excel for a team that used to copy them across by hand • A web app where staff run extraction, review flagged values and download validated data themselves ──────────────────────────── THE NO-SURPRISES BUILD (why it doesn't break on document 47) 1. Schema first. Clean output defined (Pydantic) before touching a document. 2. Rules first. Rules and regex show exactly where the hard cases are. 3. LLMs only where rules fail. If rules get you 90% of the way, I'll say so and we skip the LLM. That keeps cost per document low. 4. Layout detection only where variation needs it. 5. Your real documents as the test set. Every fix is checked against every old layout. ──────────────────────────── HOW TO START Not sure which fits? Book a 30-minute consultation ($100) and we'll decide together. 1) Diagnostic: $500 fixed, 5 business days Two versions, depending on where you are: • Pipeline Audit, if you already have an extraction script or pipeline: send up to 5 documents it gets wrong, plus access to the code or its outputs. You get every failure point found and ranked, the root cause of each, and a fix plan. • Document Diagnostic, if you're starting from documents only: send up to 5 of your messiest documents (up to 2 document types). You get which fields can be extracted reliably, which need a person to check, and a build plan with cost per document. Both include a Loom walkthrough on your own documents, a written report and a fixed quote for the work. The $500 is credited in full against the build. 2) First-Document Pilot: fixed price from $2,000 One document type live end to end, tested on your real documents against an accuracy target we agree up front. 3) Always-Working Retainer: from $1,000/month New layouts added as they appear, drift monitoring and fixes. My current retainer client has stayed for 2+ years. ──────────────────────────── QUESTIONS CLIENTS ASK "Will the AI make numbers up?" Rules handle the layouts we know. An LLM only sees new ones, and its output passes the same checks. Anything uncertain goes to a person. "What happens when a supplier changes its layout?" It's flagged, not silently wrong. The new layout becomes a test case and goes live in the next fix. "Is our data safe?" Redacted samples are fine for the diagnostic. You get a written note of where files go and when they're deleted. "Is it worth it for our volume?" The diagnostic tells you. If typing is cheaper than building, I'll say so. ──────────────────────────── WHAT CLIENTS SAY "Thanks to Subhajit's work, we are saving countless hours having to manually enter results into our own template." (Water consultancy, 2-year retainer) "His communication is clear, he takes feedback well, and he consistently finds effective solutions." "Professionalism, meticulous attention to detail, and expert handling of our project were apparent from the outset." ──────────────────────────── GOOD FIT / NOT A FIT Good fit: 100+ documents a month, layouts that vary, and data that feeds a system people rely on. Not a fit: one-off scraping of a few files, or projects where nobody can define correct output. STACK: Python · pdfplumber · PyMuPDF · OpenAI API · Pydantic · FastAPI · pandas · pytest · GitHub Actions · Docker ──────────────────────────── Message me three things: the document type, your monthly volume, and what's breaking. I'll reply within 24 hours. Top Rated Plus · 100% Job Success · $60K+ earned · 2+ years in production