I build AI-augmented systems and publish what I learned from checking them against the tasks they are meant to serve. Here are the hackathon records, product iterations, and retrospectives behind that work.
Published essays tagged Build in Public, from product experiments to lessons carried into later builds.
Public repositories and recorded checks from deadline-bound builds. The entries distinguish injected test faults from host-system bugs, and state what each result can support. Unverified competition outcomes stay unverified.
IBM Bob 2.0 Hackathon
Submitted September 25, 2026 · Award outcome not verified as of September 26, 2026
I tested Bob's tests. My own enforcement hooks once passed unit tests for days while the platform silently discarded their output; that taught me green only describes what a suite checked. At the IBM Bob 2.0 Hackathon, I built uncaught to inject controlled mistakes into pinned versions of two open-source libraries and had Bob write tests for the gaps. On quick-lru, the same failure pattern in an unseen path escaped Bob's first test. Bob revised the tests; I then challenged them with held-out mistakes and a red-team round. dot-prop had a scout-guided round but no held-out check. These were injected faults, not bugs in either released library.
A result only counts when the original suite is green, the injected mistake is valid, and a stable test catches it. Two proposed upstream tests are filed as issues; neither issue is a merged change or evidence of maintainer adoption. No competition outcome is claimed.
Read the failure that led to this build →Filed upstream while building
All Things Agentic Hackathon (Google Cloud)
Built Aug 21 – 25, 2026 · Submitted Aug 30, 2026 — judging
A monthly evidence steward for homeschool learning records. It traces an attested time floor to source rows, refuses broken coverage, and holds decisions that need more evidence. Its recorded five-reader evaluation contains 48 synthetic cases per reader; no overstated floor was attested in those 240 trials. That result describes the test set, not a guarantee about every document.
Read the build story →Build with DataHub: The Agent Hackathon
Submitted Aug 7, 2026 · Not awarded — results announced September 10, 2026; checked September 26, 2026
A DataHub agent that compares declared update cadence with a measured data timestamp. It distinguishes FRESH, STALE, UNMEASURED, and UNCOVERED, records findings in the catalog, and reads them back to check what actually landed.
Read the build story →Filed upstream while building
OpenSearch Agent Skills Hackathon
Submitted Aug 3, 2026 · Grand Prize — OpenSearch Agent Skills Hackathon, 2026
An OpenSearch investigation skill that checks the premise of a reported incident, compares each subgroup with its own baseline, and records the alternatives still untested. It keeps the investigation open when evidence is missing or the measured explanations do not account for the effect. It does not claim that logs establish causation.
Read the build story →Filed upstream while building
Qwen Cloud Hackathon — Autopilot track
Submitted Jul 7, 2026 · Results announced Aug 28, 2026 — not awarded
A hackathon prototype for maintenance quoting and negotiation rehearsal. The model drafts language, the database owns stored quote values and counts, and a person reviews the proposed action. Demonstration clients and prices are synthetic; negotiation outcomes were not measured.
Read the build story →The development diary includes experiments, mistakes, and the checks that changed what I shipped. Read the stories and the evidence behind them.