PUBLIC ENGINEERING REPOSITORY

Not one agent. A coordinated system.

The public engineering repository documents 36 specialized agent roles with scoped context, planning, approved research, and gated tool use. Read the code and follow the development state rather than taking a benchmark claim on faith.

View public repository
A precise distinction

Repository code: available

Architecture, memory, tools, training pipeline, checkpoint verification, inference runtime, and evaluation system.

Trained model: not yet

Dataset preparation, GPU training, the initial checkpoint, inference verification, and measured evaluation results are next.

01 / SYSTEM DESIGN

How the pieces fit.

The architecture is designed for explicit coordination and controlled access to context and tools.

01 / 04

Coordinate

A 36-role architecture routes work through specialized agent responsibilities instead of treating every request as a single undifferentiated prompt.

02 / 04

Retrieve context

Scoped memory retrieval supplies relevant context to the workflow without implying that every agent sees everything.

03 / 04

Plan and reason

Planning and reasoning components organize work before execution. The architecture is inspectable in the public repository.

04 / 04

Use guarded capabilities

Web research is approval-based. Tool execution is gated, and Python runs in an external sandbox.

02 / TRAINING & VALIDATION

The pipeline is not the checkpoint.

The repository includes an SFT / DPO / QLoRA training pipeline, checkpoint verification, a local inference runtime, and an evaluation system. These are infrastructure capabilities, not evidence that a proprietary model is already trained or has passed a measured evaluation.

See development progress
TRAINING WORKFLOW / REPO
01SFT / DPO / QLoRA
02Checkpoint verification
03Local inference runtime
04Evaluation system

GPU training, an initial checkpoint, inference verification, and measured evaluation results are not complete.

03 / CURRENT APP

What you can use in the app.

These product interfaces are available separately from the training roadmap.

Inspect the source.

The public repository is the most direct account of the system's implementation.

View repository