Binny C.

Your AI is only as good as the data it retrieves. Most RAG projects do not fail at the model. They fail earlier. Messy source documents. Poor chunking. Missing metadata. Stale indexes. No way to tell a good answer from a confident wrong one. I come from the data side. Since 2018 I have built pipelines that pull data out of systems never designed to share it. Now I build retrieval and AI agents on that foundation. 100% Job Success • 4.9 across 17 reviews • Shipping generative AI in production since January 2023 WHAT I BUILD - RAG over your own documents. Ingestion, cleaning, chunking, embeddings and pgvector. Hybrid keyword and semantic search. Answers that cite the source passage. - Semantic search across products, policies, tickets or contracts. People find things by meaning, not exact words. - Tool-calling agents that work through your APIs and databases. A person approves anything risky. - Document extraction. PDFs, emails and portals become structured, searchable data. WORK THAT IS LIVE TODAY - StartDeck: document automation for commercial appraisers. An OpenAI assistant answers from a BigQuery knowledge base of the client's own material. Documents are created 50 to 70% faster. StartDeck is past 1,000 installs on Google Workspace. - UK buying group: 100+ supplier sources merged into one validated catalogue of 10,000+ products. Price prep fell from five days to under one. Errors fell 95%. Bad data stops at the boundary. Ambiguous records go to a person. Retrieval needs the same discipline. - Property tax platform: 90,000+ homeowners per cycle, with 75% less manual valuation research. - UPS and FedEx parcel audits: extraction from carrier portals at scale, feeding refund disputes. "Binny is a master of his craft. He stepped into an existing n8n workflow he did not build, found the root causes quickly, and fixed the classification, routing and escalation..." STACK - Python: FastAPI, Pydantic, SQLAlchemy - Retrieval: PostgreSQL with pgvector, embeddings, hybrid search, BigQuery - LLMs: OpenAI, Anthropic, OpenRouter, structured outputs, tool calling, human-in-the-loop - Automation: n8n, webhooks, Google Apps Script - JavaScript: TypeScript, Node.js, React, Next.js - Cloud: AWS, Google Cloud, DigitalOcean, Docker HOW WE START Send me a sample of your documents and ten questions your team really asks. I reply with the retrieval design, the effort and a fixed-price first milestone. You see real answers on your own data before you commit further. Available 9:30am to 8:00pm UK time. Full UK/EU overlap and about 6 hours with US Eastern. Typical response under 4 hours.