Research & field notes
Draft essays on model serving, evaluation, and delivery. These are research and writing in progress, not published results or client claims.
Working notes from our research running 500B+ parameter open-weight mixture-of-experts models on single-workstation hardware. The draft examines expert streaming, speculative decoding, and the operational tradeoffs behind that architecture.
A proposed method for turning representative tasks and failure cases into a check on model and prompt changes. The emphasis is on pinned versions, reviewable regressions, and a clear release decision.
Notes on coordinated multi-model serving: placement, contention, and how a shared machine changes latency and capacity decisions. This follows our own systems research; no benchmark figures are claimed here.
How to specify states, exceptions, and human handoffs for software that becomes a record of work. The draft focuses on what an engineer and an operator should be able to verify together.
Write to us with a systems problem or a technical question worth examining.
Start a conversation