FOR ENGINEERS
Research.
Working notes and longer investigations — the unpolished thinking behind the shipped systems.
Scheduling LLM inference across consumer GPUs
The bottleneck isn't compute. It's what you choose to run, and when — a scheduler that picks the right backend per operation beats brute-force parallelism on hardware that was never meant to share a workload.
READ THE NOTE →Protocols for handing work between agents under partial context
READ THE NOTE →Treating models as processes in an AI-native operating system
If a model can be scheduled, it can be a citizen of the kernel, not a guest of it — the OS primitives that manage processes (priority, preemption, resource limits) apply almost directly to managing models.
READ THE NOTE →Statistical anomaly detection beats LLM-judged evals for production monitoring
READ THE NOTE →A block-based CMS is the right shape even when the only editor is you
Four client CMSes and one portfolio later: uniform blocks with dumb fields beat clever schemas every time someone (including future-you) has to actually edit content.
READ THE NOTE →Tool-calling over a voice WebSocket: keeping one contract for text and speech
READ THE NOTE →Deep repo knowledge for an on-site agent — without a vector database
For two dozen repos, pre-written docs beat embeddings: index once into a handful of LLM-authored documents (architecture, stack, per-module dives), then let the agent read titles and fetch whole docs on demand.
READ THE NOTE →