GPU & AI infrastructure release intelligence
Release intelligence is useful only when direct facts remain distinguishable from interpretation and no unreviewed machine output can quietly become a published engineering recommendation.
Independent system · product, data contracts, deployment, and evidence
DRIFT shipped a public Next.js interface and FastAPI contract over five human-reviewed insights — the whole live store, deliberately small — with grounded search and chat constrained to those inspectable records.
Raw changelogs are noisy, unstructured, and full of false positives. Teams miss critical AI infrastructure updates.
Each insight preserves the source span, direct facts, interpretation, workload relevance, confidence, severity, and one bounded check instead of flattening everything into a summary.
Typed stages extract candidate claims, a separate verifier rejects unsupported work, and drafts stay outside the trusted briefing until a human records the publication decision.
A deterministic fixture stack runs locally without an API key, while provider-backed capture is explicitly enabled, budgeted, audited, and labelled as a bounded live path.
The hosted Vercel and Railway services were intentionally retired after the hackathon to avoid ongoing costs. The source, architecture, recorded demo, and submission remain available as evidence.
I’m always interested in the constraints that make an architecture worth explaining.