Bubba Stump's Tech Emporium

Home > Jeeves & Bertie

Project 001: Jeeves & Bertie

CATEGORY: LOCAL ARTIFICIAL INTELLIGENCE   |   STATUS: EXPERIMENTAL

The idea: Can a relatively small local AI model supervise a larger model's research and drafts well enough to be useful, without burning more time and GPU cycles than it saves?

Bertie is the worker. Jeeves is the supervisor. A separate controller enforces limits on research and retries. The system is deliberately not an unlimited "just keep trying" loop.

Here is the intended control flow. This is a design diagram, not a claim that every component of the integrated system has already been deployed.

                        +--------------------+
                        |  USER'S QUESTION   |
                        +---------+----------+
                                  |
                                  v
                        +--------------------+
                        | CONTROLLER         |
                        | budgets & routing  |
                        +---------+----------+
                                  |
                                  v
                        +--------------------+
                        | BERTIE (27B)       |
                        | worker: research   |
                        | and draft          |
                        +---------+----------+
                                  |
                                  v
                        +--------------------+
                        | CHECKPOINT         |
                        | findings, source   |
                        | excerpts, origins  |
                        +---------+----------+
                                  |
                                  v
                        +--------------------+
                        | JEEVES (4B / 9B)   |
                        | supervisor review  |
                        +----+----------+----+
                             |          |
                         ACCEPT     MORE WORK
                             |          |
                             v          v
                       +----------+  +------------------+
                       |  ANSWER  |  | CONTROLLER:      |
                       |   OUT    |  | budget left?     |
                       +----------+  +--------+---------+
                                           |
                                     IF ALLOWED
                                           |
                                           +----> back to BERTIE

Fig. 1. The management insists on checking the employee's work. The controller limits how many additional trips back to Bertie are allowed.

Limits in the current design

• No more than 3 research operations in a batch.
• No more than 3 research batches total (9 operations at most).
• The first research batch can proceed without a supervisor permission step; further work is controlled.
• No more than 2 answer revisions.
• Research checkpoints preserve findings, source excerpts, and provenance so later steps can inspect the evidence.

Two computers are part of the plan. Bertie has the dedicated Linux inference box; Jeeves is destined for a separate machine. The second machine is still doing desktop duty, so this is not yet a two-host deployment.

Bertie — worker machine (in service)

Jeeves — supervisor machine (planned)

Note: The 5800X3D is earmarked for Jeeves, not already running the completed supervisor service. Its regular desktop work must move to another computer first.

Worked on: A local llama.cpp inference service, a sandboxed research/search tool, and offline comparisons of supervisor decisions across ten test scenarios. The research sandbox has been tested against local-network access restrictions.

Still being worked out: The full supervised worker loop, repeatable end-to-end measurements, and whether the extra model actually improves answer quality enough to justify its cost.

Planned next step: When a more efficient daily-use PC takes over Discord, browsing, and desktop duties, repurpose the existing Ryzen 7 5800X3D system as Jeeves' dedicated supervisor machine. Then test the real two-host workflow and measure whether supervisor overhead is worthwhile.

Project log

Return to the project index