DraftPrivate AI / Systems research
Serving 500B+ parameter mixture-of-experts models on a workstation
Expert streaming, speculative decoding, and the practical limits of single-machine inference.
Field notes
Research and working notes.
Drafts from the systems we build and the questions we study. These titles are in preparation; no articles are published here yet.
DraftPrivate AI / Infrastructure
Coordinating several models on shared hardware
Scheduling, memory pressure, and the design choices behind multi-model serving.
DraftProduct engineering / Operations
What a route-to-invoice system has to get right
Where dispatch, service records, billing, and reconciliation meet.
DraftDelivery / Practice
Acceptance criteria for software built with agents
How to make human responsibility and release decisions visible in AI-native delivery.
A conversation with the people doing the work
Have a technical problem worth studying?
Describe the system and the decision you need to make.