/blog
Field notes on IoT, web apps, AI, and whatever's currently in pieces on my bench.
Right tool for the job: why my chat engine lost the reranking race
The same model ran 9× faster after switching engines. llama.cpp SYCL won the chat workload and lost reranking, on the same Intel GPU.
2026-07-26
The models were not the only thing under evaluation. Our harness was wrong too.
What broken gates, state restoration, tiny samples, and interrupted runs taught us about evaluating LLM agents.
2026-07-15
One knowledge base, two AI jobs, and the smaller local model won one
Why the best model for answering questions was not the best model for building the knowledge those answers depend on.
2026-07-14
LLMs have Personalities - Interviewed seven models across eight lanes.
An internal agent bake-off found that serving-stack correctness, explicit briefing, and failure modes mattered more than the model leaderboard.
2026-07-13
Local LLM benchmark record: hardware, models, engines, and results
A reproducible Intel Arc Pro B60 guide with llama.cpp SYCL configuration, models, quantisations, cache behaviour, failure modes, and measured agent results.
2026-07-12
Intel Arc Pro B60 from disappointing to practical local AI Agents
How persistence, controlled testing, and a fast-moving model ecosystem turned 30-second agent turns into sub-second follow-ups.
2026-06-13