Drift Detector
Re-runs your evals nightly and wakes you only if quality actually moved. Built for Quiet Hours — build an agent that does useful work while its user is asleep. Scheduled eval runs and statistical change detection. The interface is deliberately narrow: one path, done properly, rather than a broad surface that half-works at the deadline.
Gallery
Screens
Drift Detector — main view
Drift Detector — result state
On the clock
What shipped, what didn't
Completed in time
Scheduled eval runs and statistical change detection.
Unfinished / faked
Needs a pre-existing eval set — it won't write one.
Obstacles
A dependency version mismatch cost six minutes. Pinned it and moved on.
Lessons learned
Boring, well-understood tools beat interesting ones when there's no time to debug.
Process
Prompting approach
Started from a single system prompt describing the user and their constraint, then iterated with the assistant on the Drift Detector data model before writing any interface code. Full prompt log is in the repository under /prompts.
Fairness
Disclosures
Declared by the builder at submission time, visible to judges before scoring.
Reused a personal utility file for date formatting (~40 lines), linked in the repo README.
Judging
Public feedback
Judges' private notes stay private — this is what they chose to say publicly.
Elena Voss
VP of Engineering, Harbor Compute
Solid build with a clear user in mind. The interface is doing less than it could — a second pass on hierarchy would lift this considerably.
Hana Kobayashi
Principal Research Engineer, Independent
Nice work under the clock. The unfinished list is longer than the finished one, but what shipped works and you were straight about it.
Claudia Reyes
Staff Engineer, Vantage Analytics
The idea is strong and the execution is most of the way there. The demo path broke once, which is what separates this from the top three today.
Discussion
0 comments
No comments yet.