TraceMesh icon: connected node graph on a dark rounded plate
In development — no release date

TraceMesh

A visual AI workbench for orchestration, persistent knowledge, GPU training, and robotics with Deck.

Status
In development — no release date
Platform
Not determined
License
MIT License
Released
No date announced

Overview

TraceMesh is a Windows AI workbench that connects multi-model orchestration, a persistent 3D knowledge graph, and model development. Requests, responses, artifacts, learned skills, tasks, and model handoffs remain inspectable so you can trace how a result was produced and reuse what the system learned.

NetworkMesh turns a goal into dependent tasks for specialist workers, with controlled host tools, budgets, and independent verification. Learn builds reusable skills at the context layer; a separate GPU training platform curates datasets, fine-tunes supported open models, evaluates checkpoints, and serves approved models. Closed provider models are not weight-tuned by Learn.

Deck introduces a visual robotics simulator and trainer workflow around the Unitree H2. TraceMesh helps define the session and hands it to Deck, which coordinates simulation, recordings, training, and evaluation through the connected robotics backends, with live cameras and replay in one window.

TraceMesh and Deck are both under active development. Neither is available for download, and neither has a projected release date. The capabilities below describe the current development scope, not a public release or a claim of real-robot readiness.

Development capabilities: TraceMesh and Deck

Orchestration and persistent knowledge

  • Interactive 3D graph of models, requests, responses, artifacts, tasks, and reusable SKILL nodes with provenance and lineage.
  • TraceMesh AI routing, bounded specialist consultations, model comparison, and parallel decomposition of large tasks.
  • NetworkMesh schedules dependent work, runs ready tasks, gates host tools, and checks delivered artifacts against task requirements.
  • Learn supports study, research, bounded rumination, skill transfer, and distillation; GameMesh applies orchestration to Godot and game-project work.

GPU training, evaluation, and serving

  • Curate versioned datasets and train supported Hugging Face models with LoRA on a configured host GPU; checkpoints and evaluation results stay tied to the run.
  • Describe a chat model, edit a teacher-generated specification, draft varied examples, review a random sample, and use teacher critique to flag examples for human review.
  • Compare a trained chat checkpoint against its base model on held-out examples; promotion depends on the selected acceptance threshold.
  • Whisper speech transcription and LoRA training include held-out word and character error rates. Host serving supports chat models/adapters and speech endpoints with a shared GPU budget.

Integrations and connected workflows

  • Connect configured AI providers and Hugging Face models for routing, research, teacher-assisted datasets, and host inference.
  • Signed-in workspaces use Supabase graph memory; Jeff Runtime records execution and approval context. These development integrations are separate from the website download account.
  • TraceMesh hands validated H2 session requests to Deck. Deck coordinates TraceMesh simulation and training plus NID fine-tune workflows, using configured local, SSH GPU, or dedicated fine-tune targets.
  • Model catalogs, hardware fit, datasets, checkpoints, and run artifacts connect planning to execution. Remote jobs can continue through a connection loss while Deck reconnects its status and camera mirror.

Introducing Deck: robotics simulator and trainer

  • A desktop session window for Unitree H2 simulation and training: head, follow, watcher, and wrist camera views; stage progress; training loss; closed-loop outcome rates; and failed-episode replay.
  • Record demonstrations, train camera policies or supported language-model drivers, rerun datasets on fixed seeds, and compare policy outcomes. Language-model walk drivers read simulator scene descriptions, not camera images.
  • Development scenarios cover walking and stopping around obstacles, lever doors, building doors, sedan doors and OBD-II plug insertion, sitting and waving, and compatible segments chained into continuous episodes.
  • Door workflows combine simulated and real demonstrations for GR00T fine-tuning, held-out open-loop checks, and simulator tests. Whole-body policy rollout currently covers sedan and building-door scenes; sit/wave and chained policy evaluation remain incomplete.
  • Queue sessions for configured local or remote GPU targets, inspect recordings and checkpoints, and replay failures while another stage continues. Target setup and compatible model/runtime dependencies are required.

Recent development updates

  • Expanded from context-layer Learn memory to host GPU chat and speech training, checkpoint scoring, promotion, and inference serving.
  • Added prompt-to-model dataset drafting, variation coverage, owner examples, human sample review, and a teacher critique pass.
  • Introduced TraceMesh-to-Deck session intake, remote GPU execution, camera mirroring, failure replay, and robotics dataset pipelines.
  • Extended H2 scenarios with contact-based plug insertion, building doors, chained demonstrations, whole-body policy replay checks, and updated vehicle visuals.

Who it is for

Developers, AI researchers, and robotics teams who need observable multi-model work, curated training data, reproducible evaluation, and a visual place to run and review simulation sessions.

Use it to

  • Route a complex request through TraceMesh AI with specialist consultations and a visible handoff graph.
  • Schedule a NetworkMesh goal into verified parallel tasks with tool gating and adherence scores.
  • Train reusable skills into graph memory and carry them across models without fine-tuning provider weights.
  • Build a reviewed training dataset from a model brief, fine-tune a supported open model, and compare it against its base model on held-out examples.
  • Describe an H2 robotics session in TraceMesh, review its dataset and target, then follow simulation, training, and failed-episode replay in Deck.

Available today

While TraceMesh and Deck are in development, these applications are available to download.

Explore available applications