Reserve quota before the model call, then check what keeps it safe
AlgoLens moved quota reservation ahead of its model call to close a check-then-act race. What keeps the reservation from racing, and what the test shows.
- node
- concurrency
- debugging
Notes from the field.
Production AI, debugging black boxes, offline-first mobile, and the checks that keep software honest. Short, opinionated, occasionally useful.
How I built an order bot that refuses to trust payment screenshots
Fake GCash receipts are easy to make. My chat order demo marks an order paid only when the gateway confirms the money, and code, not AI, does the math.
- payments
- llm
- n8n
JavaScript parses a date as UTC and a date-time as local
new Date("2026-07-23") is UTC, but add a time and it becomes local. The ECMAScript rule behind the split, and how to treat naive timestamps.
- javascript
- timezones
A clean QEMU run ruled out the instruction mismatch
QEMU ran the shipped llama stack with a Sandy Bridge instruction profile. It did not reproduce the crash, pointing to the next problem being elsewhere.
- testing
- ci
- qemu
OCR can read the right numbers and still choose the wrong one
GCash ads put valid peso amounts ahead of the receipt total in OCR text. Cashi checks that amount plus fee equals the total to reject ad noise.
- ocr
- parsing
- debugging
One unreadable receipt should not kill the batch
Cashi runs receipt OCR on-device with ML Kit. Its batch wrapper preserves every input path and isolates a failed image instead of dropping the run.
- flutter
- ocr
- mobile
A repository interface is not cloud sync
Cashi keeps SQLite behind repository interfaces. That gives a cloud implementation one boundary, but it does not make synchronization a one-file job.
- flutter
- architecture
- offline
OCR confidence is a routing rule, not a probability
Cashi uses a weighted completeness score and required-field checks to send uncertain receipts to review. It does not pretend to predict accuracy.
- ocr
- testing
- product
AVX2 was only the first illegal instruction
A llama.cpp wheel failed on a judge CPU without AVX2. Disabling AVX2 was not enough because the non-native build still enabled BMI2 and emitted pdep.
- llama-cpp
- debugging
- cpu
One number, no logs: five days on the AMD hackathon leaderboard
The AMD Developer Hackathon judge returned one accuracy number per Docker submission, with no logs or stderr. Five days of debugging through that single score.
- ai
- hackathon
- debugging
- llm
Adjectives aren't a spec
I judged an image pipeline for months by whether the output looked right, until the metric that said PASS started quietly disagreeing with the pixels.
- engineering
- quality
- testing
- image-processing