🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Pocketful: All 4 Stages passed
Pocketful tests whether a reusable team of coding agents can build and verify a payments service through BAND. An Architect coordinates a Builder, an independent Verifier and a Spec Auditor, using generic seat mandates for ownership, handoffs and review. The result is four independently buildable Docker services. Stage 1: payments and settlements. Stage 2: a browser wallet with payment authorizations, held funds, expiry and captures. Stage 3: statements, historical balances (as_of, known_at), frozen pagination snapshots and idempotent payment corrections with historical overdraft checks. Stage 4: receiver refunds and atomic operator batch corrections across settlements. Each stage keeps earlier code unchanged in its own folder and Git history. Fresh-clone isolated harness runs passed all 147 Stage 1, 35 Stage 2, 6 Stage 3 and 5 Stage 4 checks, with earlier stages re-passing cumulatively. Both independent reviewers approved each stage's exact revision; for Stage 3 they first found three issues, which the Builder fixed. We include source, mandates, the factory description and the unchanged full BAND export. This was a supervised practice run with human execution approvals, host debugging and UI checks, not a fully autonomous judged run. No paid fallback was used, and we did not measure per-seat cost. When the Codex seats hit their usage limit before Stage 4, the owner approved substituting Claude Builder and Verifier seats (disclosed); both reviewers approved Stage 4 at 203e448.
6 Oct 2026
GroundMesh Inspect: Vision-Guided Robot
GroundMesh Inspect turns robotic manipulation from a one-shot prediction problem into a closed-loop, self-correcting system. Instead of asking a single model to perform an entire task and hoping it succeeds, our robot uses three specialized ACT policies for the core manipulation stages—moving an object to a known center position, placing it into the correct bin, and placing rejected objects into a bad bin. A visual inspection layer powered by Anomalib continuously evaluates the robot’s state and the result of each action. If an object is misplaced, the grasp is unstable, or the observed outcome does not match the expected state, the system can reject the result, select the appropriate recovery policy, and retry rather than blindly continuing. We combine this with OpenVINO-optimized inference to reduce latency and jitter on Intel edge hardware, turning separately trained policies and vision models into one real-time decision pipeline. The result is a practical architecture for reliable robotics: act → observe → verify → recover → continue. Rather than only demonstrating that a robot can complete a task under ideal conditions, GroundMesh Inspect shows how learned policies, visual quality control, and explicit recovery logic can work together to make physical AI more dependable in the real world.
16 Sep 2026
California State Spending Data Deep Dive
RealWork is a public-spending intelligence pipeline built on live web data from Bright Data. We ingested 100,000+ LA City payment records, cross-referenced 500 California nonprofit grant recipients against their IRS Form 990 filings, and scanned 50,000 state purchase orders for threshold-structuring patterns. Then we built a second capability: a PDF extraction engine that turns LA County's 1,652 unqueryable scanned homeless-service invoices into a live, public, structured dataset. What Bright Data enables: CA Secretary of State blocks ordinary scrapers. ProPublica's 990 archive requires bypass. The state's procurement dashboard is a JS-rendered SPA. Web Unlocker, SERP API, and Scraping Browser are what make cross-source verification possible at scale. What the pipeline found: 12 Cal Fire purchase orders at exactly $49,950 — $50 below the competitive-bidding threshold — signed by a single procurement officer in a 17-day window. 36 nonprofit grant recipients with officer compensation anomalies flagged from public 990 filings. And a live extraction of $3M+ in homeless-service billing data that LA County publishes as scanned PDFs nobody can query. The extraction pipeline: Gemini 2.5 Flash reads scanned PDFs natively — no OCR stack — and returns structured JSON (vendor, date, billed amount, deliverables). The ledger is public on GitHub and growing as the pipeline runs. Every finding cites public records. No fraud conclusions — anomalies warranting investigation by oversight bodies with subpoena authority.
31 May 2026
