Engineering write-ups from production AI systems.
Long-form posts on what we actually measured — voice agent latency broken into seven segments, on-prem LLM evaluation ceilings, retrieval design decisions. Numbers copied verbatim from the runs, not the pitch deck.
AI voice agents for restaurants: why the first turn feels slow (and how I fixed it)
In a real-time voice pipeline (SIP + streaming STT + LLM + streaming TTS), the first response after "Pickup" was consistently 2 s slower than every later turn. STT and TTS were fine. The culprit was a cold prompt cache on the LLM. A warmup fired in parallel with the greeting cut first-turn latency by ~50% for about $0.02 per call.
Automating medical coding on on-prem LLMs: what we measured before building anything
An engineering study on ICD-10-GM 2026 for a German inpatient hospital, fully on-premise. Before choosing a model we measured the retrieval ceiling — 0.854. The naive "embed the note, find a similar code" architecture tops out at 3.9%. The standard way of parsing the official BfArM catalogue silently loses 27% of the code space.
Bring us a system you can't measure yet.
One 20-minute call. Voice agent latency you can't explain, a retrieval ceiling you haven't measured, an on-prem constraint that changes the architecture — we'll tell you honestly whether we can help, and what the first two weeks would look like.