Return to front page Archive
hmntrjpl-labs

AutoResearch Lab

An operator-facing lab for two job types: GPU-backed `karpathy/autoresearch` experiments and a new Workflow Studio for repo analysis, patch planning, and review routed through free OpenRouter models.

Separate control plane

Research jobs stay off the static site.

The public site remains a readable front page. This lab page talks to a separate control API, which then dispatches work to a GPU worker or a repo-analysis workflow runner. AutoResearch keeps its pinned `karpathy/autoresearch` target, while Workflow Studio accepts a repo URL and produces staged artifacts instead of writing directly to code.

Operator lanes

AutoResearch: NVIDIA-led council for experiment loops against the pinned autoresearch repo.

Workflow Studio: Nemotron for planning, Qwen Coder for backend patches, StepFun for frontend/UI, and Qwen Next for review.

Current snapshot: No published runs yet

Admin-only launch Static site, separate API Patch proposals only
Pinned external dependency

AutoResearch Desk

Use this when you want a GPU worker to pursue a new experiment against the pinned `karpathy/autoresearch` repo.

Launches require the separate lab API to be configured.

Latest experiments

AutoResearch Runs

Completed and in-flight GPU experiments, newest first.

No experiments yet.

CCG-inspired orchestration

Workflow Studio

Submit a repo, choose an operator task, and let the staged council produce a plan, backend patch, frontend patch, and review verdict without granting direct write access.

Workflow Studio also requires the separate lab API.

Latest workflow runs

Workflow Studio Rail

Recent repo-analysis and patch-proposal runs, surfaced as staged artifacts rather than direct edits.

No workflow runs yet.

Run detail

Inspect a run

This panel surfaces the generated task brief, model attempts, artifacts, and stage-by-stage output. Workflow Studio runs also show the planner, backend patch, frontend patch, and review verdict in order.

Select an experiment to inspect its program, status, and artifacts.