Operating model and governance
The failure I keep seeing is not that the model was wrong. A demo worked, someone called it production, and nobody can say who was allowed to ship it. I treat the path from spec to release as the product, not the model sitting in the middle of it.
In the company I co-founded, humans hold named roles at spec, review, and release. Work does not move until those gates are cleared, and a release approval is recorded. We use the same discipline on language. A capability is Built, Deployed, Wired, Demonstrated, or Validated. If it has not earned a label, we do not claim it.
- Named human roles for spec, review, and release
- Gated delivery from spec clarification through recorded human release approval
- Capability claim standard: Built, Deployed, Wired, Demonstrated, Validated
Evaluation and decision discipline
Most evaluation starts after people already know the answer they want. I write the decision bar down and sign it before anyone begins the research. That keeps the later argument honest, because the standard cannot be moved to rescue a favorite idea.
That is the design behind the Foundry. Each idea goes through a fixed pipeline: multi-agent research, a Bear Memo that argues the other side, a weighted scorecard, and a recorded verdict. I ran ten ideas through it. The point was not to generate more work. It was to make go / no-go a matter of record.
- Preregistered decision bars, signed before research
- Multi-agent research, Bear Memos, weighted scorecards, recorded verdicts
- Foundry run across 10 ideas with PURSUE / ITERATE / PARK / PASS outcomes
Model strategy and AI economics
Inference cost usually shows up as an engineering leftover. It is not. If a workflow is only affordable on a promotional API rate, you do not have an operating model. You have a discount.
We run local inference on Qwen and Gemma on a 512 GB Mac Studio, behind a role-based router. I contributed to the model selection and the routing rules, and I co-administer the machine fleet. The standing rule is to prefer subscription and local compute over metered API unless the job actually requires the API.
- Local stack: Qwen and Gemma on a 512 GB Mac Studio, role-based router
- Contributed to model selection, routing rules, and the local-over-API cost rule
- Co-administer the machine fleet
Enterprise buying and adoption
Regulated buyers are not shopping for a model. They are shopping for a path they can defend: CISO, GRC, technical evaluation, then procurement. I have sold through that sequence in federal cybersecurity, including twelve-month cycles that started as paid discovery and ended as standing programs.
The case I write is the one those rooms can carry internally. Risk reduction and cloud economics survive a review. Model novelty does not.
- How CISOs, GRC teams, technical evaluators, and procurement decide, and in what order
- Twelve-month federal cycles; paid discovery converted into programs
- Business cases framed on risk reduction and cloud economics
