Three jobs. The same platform underneath.
Start with what the assistant has to do: answer from approved knowledge, carry a long-running thread, or run privately on hardware you control. The industry changes; the pattern does not.
Start with the work, not the industry.
The same patterns recur in finance, healthcare, manufacturing, public-sector work, and support. Pick the job that looks like yours.
Answer from approved files.
Retrieve the passage that answers, cite it, and use Governance when some topics must be refused altogether.
Documents + Governance →Keep the history without resending it.
Give each step the useful part of the backlog instead of paying every agent to reread everything.
Context →Run the model where the data lives.
Use compressed weights on existing GPUs, in your VPC or fully air-gapped.
Models →Three concrete versions of the same boundary.
Each example shows what the assistant may answer from, what it must refuse, and where it runs.
Bring us the documents, and the topic you can’t let it discuss.
We’ll ground it, lock the topic, try to get past it, and show you the record. Thirty minutes, an engineer, no deck.