Systems
Architecture before ornament
Making a system look impressive is easy. The hard part is deciding what has to keep working when it's under load.
Short notes on systems that have to ship, and stay up.
Systems
Making a system look impressive is easy. The hard part is deciding what has to keep working when it's under load.
AI
A model on its own isn't a product. Evals, fallbacks and a clear owner decide whether AI belongs in the system.
Delivery
Shipping is only the first test. What matters is whether the system still works after launch week.
Data
A dashboard is the easy part. The hard part is a pipeline that stays accurate when the source systems change.
Systems
A good engineer should be able to run your system without asking the person who built it. That starts with knowing what owns each piece of state.
Each one is an open-source repo we installed with no API keys, measured, and wrote up with the numbers.
We wrapped strands-decider as an allow, deny or escalate merge gate. Keyless run: 12 tests passed, deny exit 1 at 0.91, escalate exit 3 at 0.42 under a 0.8 threshold.
SkilledDragonCorp/minilab-decider-merge-gate
We wrote a default-deny Dogwood policy for Strands Box. Keyless: 3 of 3 cargo tests over 26 commands; git status allowed, git push origin main denied with exit 1.
SkilledDragonCorp/minilab-box-contain-merge
We pinned Open Code Review v1.13.0 behind a fail-closed PR gate. Keyless: 8 of 8 tests and 8 of 8 lab CI jobs passed. Without a provider key the findings come from our own scanners.
SkilledDragonCorp/minilab-ocr-scion-review
We pinned planning-with-files and tested recovery from disk and its completion gate. Keyless: 6 of 6 pytest tests passed, and upstream pytest and vitest ran green in CI.
SkilledDragonCorp/minilab-planning-with-files-gate