Neurarch documentation

Neurarch turns a neural network into a typed graph that an AI agent can edit, runs structural checks on it, and hands you the PyTorch you would have written. This page is the short version of how to use it.

What it is First caught bug Start a model Design to deploy The agent Structural checks Export and train Accounts and plans Import a model Feature reference FAQ

Reading this with an AI agent? These docs are also served as plain markdown: start at /llms.txt.

What it is

Every model in Neurarch is a typed graph: nodes are layers or blocks, edges carry tensor shapes. Because the graph is typed, the app can check it for real bugs (shape mismatches, broken residuals, attention head-dim errors) before you ever spend GPU time, and it can generate framework code directly from the structure.

It is not a drawing tool. The diagram is a side effect of the structure, not the point.

Your first caught bug (60 seconds)

The fastest way to understand Neurarch is to watch a check fire. The link below plants a classic bug (attention with embedDim 768 and numHeads 10; 768 / 10 is not an integer) and shows you the catch:

1Open the app with the planted bug. No account needed; the graph loads entirely in your browser.
2The pre-flight report opens on the finding: a blocking shape check, embedDim % numHeads != 0, before any GPU is involved.
3Close the report. The same finding lives on the canvas as a red dot on the attention layer: hover it for the explanation and the suggested fix, click it to jump to the exact finding in the Architecture Advisor.
4Fix the head count (12 works) and the health badge in the stats bar turns green.

Shipped instead, this bug is an AssertionError raised when the module is constructed on the training node, after the job was queued and the dataset downloaded. More planted bugs, and the measured evidence that the checks track reality (264 graphs cross-tabulated against real PyTorch): caught before the GPU bill.

Start a model

The prompt box on the home screen is the primary path: describe what you want to build and hit Design and verify; the agent assembles the architecture on the canvas. The four cards under it cover the other ways in:

Click any card for the details.

Pick by taskWe suggest an architecture for your task Pick by modelStart from a known architecture Browse the gallery1M+ models, papers, community designs Check my codePaste PyTorch, get the bugs

You can also import from an arXiv paper or any article URL, from a sketch or screenshot, or from a HuggingFace model, or start from a blank canvas.

Design to deploy

The top bar follows the same path a model takes from idea to production:

1Problem: describe the task, pick data and output types, and let the agent build the architecture.
2Code: view and export the model as PyTorch (and other frameworks).
3Handoff: size the model, estimate cost, and pick a training backend (Modal, Replicate, RunPod, and others).
4Train: run training on a real GPU or in simulation, with live metrics.
5Deploy: export a FastAPI server, Dockerfile, and a full project bundle.

You connect your real training data at the Train step, not while designing. Designing only needs the shape of your data, not the data itself.

The agent

The agent edits the graph through structured actions, not free text. Before any change touches your model it runs a guardrail pipeline (impact analysis, parameter-explosion check, cycle and orphan detection, and shape inference), so a suggested edit that would break the model is caught and explained first. The agent requires sign-in.

Structural checks

Because the graph is typed, Neurarch can flag problems statically: tensor-shape mismatches between layers, attention where embedDim % numHeads is wrong, grouped-query attention head ratios, merge operations whose parents do not match, and residual or skip connections that do not line up. These show up on the canvas before you train.

Each flagged layer gets a colored dot (red error, amber warning, blue note). Hover it for the finding, a suggested fix, and a one-click Fix with AI; click it to open the Architecture Advisor at that exact finding. The stats bar at the top of the canvas shows a health badge ("2 warnings", or a green check when clean) you can click to review everything at once.

Export and train

From Studio and the File menu you can export: PyTorch nn.Module, a runnable Jupyter notebook, a TikZ figure for papers, a publication-style architecture figure (SVG or PNG, pastel and print-friendly), a model card, an architecture report, slides, a deploy bundle (FastAPI + Docker), or a full project zip. Training runs through Modal or in simulation; you can also export a training project and run it on your own infrastructure.

Accounts and plans

Designing, importing, and most exports are free and need no account. The agent needs sign-in. Heavier features (some LLM-backed exports, real GPU training, larger quotas) are on the paid tiers. See pricing in the app for current limits.

Bring an existing model in

If you already have an architecture, you do not have to rebuild it by hand:

From arXiv or any article URL: paste an arXiv link or ID, or a blog/article URL, and we parse the architecture from the page.
From PyTorch code: paste an nn.Module and we reconstruct the graph.
From HuggingFace: load a model by id; we fetch its config and build the matching graph, covering transformers, vision, diffusion, and ASR families (Whisper, Conformer-CTC / Parakeet). Tensor shapes are inferred on import so the graph shows real shapes immediately.
From a figure: import an architecture diagram image and we extract the structure.
Cross-validate paper against code: check that a paper's described architecture matches a code implementation.

Every import is verified before it loads, so a malformed or degenerate architecture is rejected with a reason rather than silently loaded.

Feature reference

Click any card for the details.

Canvas & editingBuild and inspect the graph BlocksReusable, repeatable sub-graphs Pattern detectionKeep large models readable AgentEdit by describing the change Scaling & sizingParams, FLOPs, target budgets TrainingReal GPU or simulation Cost & handoffEstimate cost, pick a backend DeployShip the model ExportsCode, figures, docs Sharing & embeddingLinks, iframes, gallery Community galleryBrowse and fork designs Keyboard-firstShortcuts for most actions

FAQ

Is my data uploaded?

When you drop a CSV at the design step we read only the column names and types to size the model; the file stays in your browser. Your training data is connected later, at the Train step.

Do I need to know PyTorch?

No. You can build with the task wizard, templates, or the agent, then read or export the generated PyTorch when you want it.

Can I bring my own API key?

Yes. If you supply your own model provider key it is kept in your browser and used directly; it is never sent to our servers.

Where do I report a problem or ask for help?

Use Contact in the app's "?" menu.

This page is a living summary and will grow. For anything not covered here, reach us through Contact in the app.