Latest writing
Three vector stores, three shapes of work. A field-tested rubric for picking one you will still be happy with six months after launch.
Running vision models on Jetson devices is a different sport than inference on a GPU cluster. A field checklist for quantization, runtime choice, and the thermal realities nobody puts in the datasheet.
Prompts behave like code that nobody wrote tests for. The lightweight harness we use to version, evaluate, and roll back prompts without stalling the product team.
The last 10% of moving a model to production is 90% of the work. A walkthrough of the refactors that hand a research notebook off to the oncall rotation without regrets.