Dhawal M.

I build AI agents that do a real job in your business and don't make things up. Every answer cites your own documents or data, it says "I don't know" when the answer isn't there, and you get a test report on your real questions before launch. WHERE THIS HELPS - Your support or sales team answers the same product, policy or spec questions every day. - Your chatbot or RAG app sounds confident and is sometimes wrong. - It worked in the demo and got worse in production, or the person who built it left. - You want an agent to act in your CRM, inbox, sheets or accounting software, without it quietly breaking things. WHAT I BUILD - AI agents and assistants over your docs, CRM, email, spreadsheets and internal tools: tool calling, multi-step workflows, structured outputs. - Knowledge assistants (RAG) that handle the messy parts: tables, scanned PDFs, versions that contradict each other. - Integrations with the systems you already run, through APIs, or through validated exports when there is no API. - Built with Python, the Claude and OpenAI APIs, SQL and Postgres, LangChain when it helps, n8n or Zapier when that is the simpler answer. HOW I MAKE IT TRUSTWORTHY - Cite or refuse: every answer points to the page or record it came from. No source, no answer. - Evals from your real cases: a gold set of 20-50 real questions, scored automatically, re-run on every prompt or model change, so a regression shows up before your customers see it. - Read before write: agents that change data go through hard checks and a human approval step until they have earned autonomy. - Cost and quality side by side: I test cheaper models against your bar, so you pay for the accuracy you need and no more. - Monitoring for silent failures: you hear about a run that didn't happen or came back empty, instead of finding out a week later. WHAT YOU GET A working agent your team uses, a plain-English test report (what it gets right, what it refuses, where it fails, cost per run), and documentation so it runs without me. PROOF YOU CAN READ (see Portfolio) - tally-aiagent: lets an AI agent post to accounting software without corrupting the books. It refuses names the ledger doesn't know and reads every entry back to verify it. Running in production. - doceval: my open-source eval harness for AI extraction. Field-level accuracy, a failure taxonomy and cost per run. On PyPI. - Past Upwork automation work for repeat clients, reviewed below. HOW I WORK I ask about the pain before the tech, start with a small fixed-price milestone so you can judge the work, and send you the test results, not just the demo. Not a fit: rating, annotation or AI-trainer gigs.