Participated in public demo 2603 with CARA

The problem [00:00:10] The EU AI Act requires all AI systems active in the EU to be classified before August 2026, miss the deadline and the penalty is up to €35M or 7% of global turnover. Most organizations were managing this with spreadsheets, email threads, and manual sign-offs. What they built [00:01:20] David built CARA, an end-to-end compliance pipeline that takes a submitted AI use case, screens it for completeness before any model sees it, and retrieves only the most relevant sections of EU AI Act text to ground the classification. Two independent AI models classify the case in parallel without shared context, and a human reviewer makes the final, binding call, producing a signed, tamper-evident audit record. He iterated through three sprints, a no-code proof of concept, a Maker-Checker pipeline exposed by an incomplete legal corpus, then a full pipeline grounded in retrieved legal text, fixing a corpus gap, a silent retrieval bug, and a citation-normalization bug along the way. What happened [00:01:20] Full pipeline runs end-to-end for under $0.06 per case. [00:06:20] David ran 10 priority evaluation cases across banking, healthcare, HR, pharma, and critical infrastructure, spanning the full risk spectrum: 5 consistent between the two models, 4 divergent, and 1 blocked by the confidence gate before reaching the models at all. One case (a "Nudge Savings Assistant") is flagged as a likely miss, both models called it minimal risk, but David believes it's Article 5 manipulation, which he cites as evidence the system needs governance-professional validation before it can be trusted. The ask [00:07:25] David is looking to connect with regulated enterprises with AI use cases in flight ahead of the August 2026 deadline, and especially with compliance, legal, and privacy professionals who can stress-test CARA's classifications.