Insights
Notes from production
3 September 2026
Taking the system in-house
Handover is a success case — code, IaC, runbooks and evaluation suites are yours from day one.
28 May 2026
Inference cost is a product decision
Model choice, prompt design, retrieval scope and caching shape the monthly bill as much as usage — and product owners should own that trade-off.
14 May 2026
LLMOps after the notebook
Notebooks prove feasibility; LLMOps proves you can change prompts, models and corpora without breaking production.
19 March 2026
What a two-week discovery actually produces
Ten working days, three people, one readout — and a discovery pack your board can act on.
5 March 2026
Most enterprise AI never leaves the demo
Pilots impress steering committees and stall in production for predictable reasons — not because the model was wrong.
25 June 2026
Give AI an API it can trust
Agents fail on ambiguous tools, silent errors and missing idempotency — reliable APIs are the real integration work.
11 June 2026
Governance is not a phase
Logging, oversight and evaluation belong in sprint one — "phase two" is how audit findings happen.
9 July 2026
Adoption is part of the build
Training, comms and feedback loops belong in the definition of done — not a slide after go-live.
6 August 2026
Who is awake at 3am
Production AI needs runbooks, on-call and clear ownership — the question every steering committee should ask before go-live.
30 April 2026
The data work pilots skip
Retrieval, lineage and refresh schedules determine whether AI stays accurate — and pilots treat them as tomorrow's problem.
16 April 2026
Agents that act, not assistants that chat
The difference is tool use, supervision and workflow — not a bigger chat window.
2 April 2026
Evaluation before the first user
If you cannot test it, you cannot ship it — evaluation harnesses belong in sprint one, not after launch.
Next step
Start with two weeks.
A fixed-fee discovery gives you a ranked use-case portfolio, a target architecture, a cost model and a build proposal you can take to your board. If we don't find a case worth building, we tell you.