Participated in public demo 2605 with Vault

The problem [00:00:00] Every adult juggles personal-record documents scattered across emails and logins, car insurance, payslips, passports, eye prescriptions, and spends "10s of minutes" each time just tracking down one piece of information. What they built [00:01:46] Vault is an AI agent (built on AWS Bedrock) that identifies a document's type on upload, looks up a per-domain schema via a tool-use "registry" pattern, and extracts structured fields with a human-in-the-loop confirmation step before filing. Once filed, documents become queryable in natural language, with every answer citing its source document ID. The system emits an event for every action (extraction, query, feedback), giving built-in observability, per-request token/cost tracking, and, per Bhavesh's reflections, the same event stream doubling as an eval substrate. What happened [00:06:28] Built and deployed a working system that files documents across multiple domains (motor insurance, eye prescriptions, share holdings) and answers natural-language queries with sourced, traceable results. Added an eval suite catching real bugs, two test cases initially failed due to a schema-registry setup issue, since fixed, giving confidence to iterate on prompts/models going forward. The ask [00:07:49] Happy to talk architecture, what's next, and where this could be useful, connect via LinkedIn.