Blog
Building better models from what happens in the field.
From raw counters to product telemetry
A guide to telemetry for an on-device LLM built with llama.cpp — turning llama_perf_context_data counters into throughput, time to first token, and failure rate across your real device fleet.
June 30, 2026Building a car recognition application (part 1)
Starting with a cloud LLM, going to a trained successor
June 12, 2026Speech-to-tool pipeline performance measured
We benchmarked on-device speech-to-tool latency across iPhone generations. The fastest two-step Whisper Tiny plus LFM2 350M path was 341.5 ms median, while direct audio-to-tool on iPhone 16 Pro landed at 1.36 seconds.
April 21, 2026Your ML telemetry should live in your data stack, not ours
Every ML observability tool eventually ships an export feature. That is a symptom, not a solution. Here is what the architecture looks like when you get it right from the start.
April 12, 2026The bug that looks like healthy infrastructure
A PCB inspection line runs eight healthy Jetson units. Latency is flat. GPU utilization is normal. Then the customer returns start.