Case study
Private AI
I run the models next to the notes and the code. Both sit behind tools with a contract. That agreement sits next to rules that decide what may leave.
Designed and operate it · 2024 – present · Running
Before
A new chat is a paste into someone else's system. That host starts the work over, and the notes leave the building.
Now
A system I run searches the notes and uses tools. Those capabilities sit behind rules that decide what leaves.
Context
The project's information used to leave the building as a paste into someone else's system. That host then held the collection, and anyone on that side could see it. The access is what I took back onto a box I run. My machine still has to make the collection usable. I control the project's information and who can see it, with a picture precise enough to change the right files while sending less of that collection out.
Constraints
- Private notes stay here.
- A new session still has the project.
- A tool names what it takes and how it fails.
- Blast radius stays small on purpose.
What I built
- 1.Notes — Linked notes stay here, searchable.
- 2.Find — That search returns the right notes, plus the right slice of the code.
- 3.Model — With that piece I run the model on this machine.
- 4.Tools — Fourteen-plus tools sit on that box, each with a way to fail and a shape.
- 5.Rules — That form is written down as what it may do, and what may leave.
The leave decision is the last check. That review sits around the model with finding the right notes and giving it tools.
Hard problems
What it can see
Every chat session that starts over is a blank slate that takes the brief as a fresh paste. I put centralized scoping and designs in a vault so the agent has documents to see. I built tools that emit a map of the code at more than one depth. My agent sees domain knowledge, research, shared notes, plus the code mapped at the project, the data, and the blast radius.
What it can do
A new ask used to mean a new way of working. I defined the workflow I needed the agents to follow. Skills inside that walk say what the task is and how to do it precisely. The structure around those skills is the consistency. Agents follow that walk from the intent in the documents to the defined spec. The map tools turn the picture they already have into a precise edit that carries the goal, the design, plus the patterns we like. Those habits keep the work the same kind of flow, and each project is richer than the last because the agents gain from cross-project learning.
What it may not send out
I ask what a hostile user could make it do, then I shrink that. That reduction is a set of rules that check the action before it runs. I can replay a session after the action fires and see what it did under that review. That review is what a new tool faces in the pipeline. We control the data and who can touch it, and a precise tool call sends less of it out to get the answer back.
Outcome
- Where it runs
- A system I operate
- See
- The right notes, plus a map of the code
- Do
- 14+ tools. Blast radius before an edit.
- Leave
- Rules decide what leaves.
Close
- What this proves
- I can put a model next to private work without handing that project to someone else's system.
- What I would keep
- Find the slice. That piece is what a contracted tool may touch, and leaving stays a rule on the way out.
- What is still open
- A new tool still has to take the same checks. Those reviews live on a machine that is mine to keep running.