DRAFTPRIVATE AI / SYSTEMS
Running 500B+ parameter open-weight MoE models on a single workstation
Expert streaming, speculative decoding, and the practical boundary between available hardware and useful inference.
NotesACYCLIC / NOTES
Working notes
from the field.
Research and implementation questions we are developing. Entries below are drafts, not published findings.
Technical notes, delivery letters, and applied research.
We write about the engineering decisions that change whether a system can be trusted in use.
DRAFTPRIVATE AI / SERVING
Coordinating multiple models on shared hardware
Scheduling, contention, and the operational choices behind a multi-model service.
DRAFTDELIVERY / ACCEPTANCE
Acceptance criteria for software that touches the books
How to specify reconciliation behavior, exceptions, and review points before implementation.
DRAFTFIELD ENGINEERING
What an embedded engineer needs to learn in the first week
Finding the real workflow, the record of truth, and the person who can accept the result.
NEXT STEP
Discuss an engagement Have a problem worth writing through?
Tell us the system, the constraint, and the deadline. We will tell you if we can take responsibility for the result.