Document Intelligence & AI Search — turn your documents into structured, searchable data
About this Gig
Stop paying people to read documents and retype what is in them. I build extraction and search systems that turn contracts, invoices, transcripts, and reports into structured data your team can filter, query, and sync. This is my core specialism. At Novis.AI I built Smart Tables, an LLM extraction platform that pulls structured properties out of documents, transcripts, and websites and auto-generates the columns that fit whatever it is reading. At Zeplyn AI I built RAG search over a wealth platform's own records and unified its Salesforce and Wealthbox integrations. WHO THIS IS FOR The answer exists in a document nobody can find. The same information gets typed into three different systems. Someone spends their week copying fields out of PDFs into a spreadsheet. WHAT YOU GET • EXTRACTION PIPELINE — structured fields pulled out of PDFs, scans, emails, spreadsheets, and transcripts. Handles messy real-world inputs: bad scans, inconsistent layouts, multiple languages. • AI SEARCH OVER YOUR OWN CONTENT (RAG) — ask a question in plain language, get an answer with the source passage cited, so your team can verify rather than trust blindly. • A REVIEW INTERFACE — a table where people can see, filter, correct, and approve what was extracted. Human-in-the-loop where judgement is genuinely needed. • SYNC TO YOUR SYSTEMS — results pushed into your CRM, ERP, database, or warehouse, so data is entered once and lands everywhere it belongs. • AN EVAL SUITE — measured accuracy per field on your real documents, so you know exactly where it is reliable and where a human should still look. I do not ship extraction without this. • Full source code, deployment, and documentation. Yours to keep and extend. COMMON BUILDS • Invoice and purchase-order data entry • Contract clause extraction and obligation tracking • Onboarding and KYC document processing • Meeting and call transcript summarisation with action items • Internal knowledge search over policies, SOPs, and past projects • Competitive and market research pulled from the web into a table HOW IT RUNS Week 1 — scoping: your documents, your fields, your accuracy bar, plus a proof of concept on real samples. Weeks 2–4 — build in weekly increments you can use, each with a demo. Week 5 — integration, eval report, handover and training. Fixed price per milestone, no hourly billing. WHY ME OpenAI contract engineer building agent evaluation infrastructure. Founding engineer at Novis.AI, Zeplyn AI, and Flasho (YC W20). ML background in OCR and document parsing from SpringWorks. More at https://pravaahtech.ai AFTER LAUNCH An optional monthly retainer covers monitoring, model updates, eval maintenance, and new document types as they appear. Cancel any month. Never required. Message me with the document type and roughly how many you process, and I will tell you honestly whether this pays for itself.
Requirements
To start, I need five things: 1. SAMPLE DOCUMENTS. 20–50 real files of each type you want processed — the messy ones included. Bad scans, odd layouts, and edge cases are the most useful thing you can send me; a folder of perfect examples will produce a system that breaks in week one. Redact sensitive fields, or we sign an NDA first — whichever you prefer. 2. THE FIELDS YOU WANT OUT. A list of what should be extracted from each document type — invoice number, date, line items, party names, renewal clause, and so on. A spreadsheet with the columns you wish were already filled is the ideal format. 3. 10–20 DOCUMENTS ALREADY FILLED IN CORRECTLY. This is the ground truth for the eval suite. It is how we measure accuracy per field instead of arguing about whether it "feels" accurate, and it is the single most valuable thing you can give me. 4. YOUR ACCURACY BAR AND WHAT HAPPENS BELOW IT. What percentage is good enough to be useful, and what should the system do when it is unsure — flag for human review, reject, or take a best guess? This varies enormously by use case, and it changes the design. 5. WHERE THE RESULTS GO. A CSV or dashboard, your database, or synced into a CRM/ERP — with API docs or credentials when we reach that step. Also, where the documents arrive from: email inbox, upload, shared drive, or an existing system. HELPFUL: rough monthly document volume, the languages involved, and who will review flagged results day to day. If you only have items 1 and 2 right now, that is fine — message me and the scoping week produces the rest.
