.png&w=256&q=75)
1
1
India
10+ years of experience
iOS Engineer and Tech Lead, Agentic AI and ML research (IIT Jodhpur MTech). Based in Mumbai, India. 12 years of experience. Skills and interests: agentic AI, LLM agents, retrieval and RAG, MCP, Python, TypeScript, Next.js, iOS and Swift. Building an AI agent that runs a real business process end to end on Evolus, powered by an open model served with vLLM on AMD Developer Cloud. First use case: a support desk that answers from company documents with sources and hands off to a person when unsure.

Every growing company gets the same long security questionnaires from customers, and answering them eats days. The answers live in policy documents, but the person filling the form usually writes from memory, and some of what gets written cannot actually be proven. That is a real risk when a customer audits you later. ProofDesk drafts each answer only from the company's own documents, with citations. Before an answer can ship, every claim in it is checked against the passage it cites. A decision layer scores each claim first: clearly supported claims ship, clearly unsupported ones are cut, and borderline cases go to a larger checker model. Anything the checker cannot settle comes back as uncertain and lands in a human review queue with the full evidence trail. Questions the documents cannot answer become hand-offs, not guesses. The sample run in the demo uses a fictional company, KiteMetric. One question asks about ISO 27001 certification. The documents only support SOC 2, so the plausible sounding ISO claim is blocked and flagged for a human. That catch is the product in one moment. Built on AMD: drafting and checking run on open models served by AMD Token Factory (DeepSeek-V4.1-Flash for drafting, Gemma-4-31B-IT as checker), and the decision layer that scores every claim first is Jev, a dedicated decision model. The finished answer sheet is dispatched into an Evolus workflow through the Evolus API, where the business process continues. ProofDesk is a containerised FastAPI app with a public repository and a live demo.
18 Oct 2026