Toró D'Extoques
"Do you know what's in your own stockroom?" AI inside the registration flow a company already has: in four days we sanitized a 23,554-item industrial materials base — 57% of it got at least one proposal, 16,721 in total, including 1,124 merges and 6,083 NCM fixes — and a chat agent over MCP registers new items to the PDM standard.
Challenge 01 of Hackathon VX gave us a synthetic master base of 23,554 industrial materials, a PDM registration standard and 520 pending registration orders: duplicated codes, free-text descriptions, wrong units and inconsistent fiscal codes (NCM). Five of us shipped a working MVP, a market analysis and a five-minute pitch between 5 and 8 November 2026 — 170+ merged PRs on a linear history.
I worked across the stack: the PDM validator, attribute extraction, identity rules and merge proposals in the batch; the MCP server, its tools and the hybrid similar-material search (full-text plus optional embeddings, measured on 202 declared duplicate pairs); the database schema, migrations and per-service SQL credentials; and the AWS workshop deployment with CloudFormation, Cognito personas, Fargate batch execution and GitHub OIDC deploys. Every algorithm that could disappoint sits behind a swappable, measured repository, and every quality number is labeled as a development measurement, not an independent one.