New · Sep 2026 A plan on every model pull request: params, cost, GPU fit, and a red check when it will not run. Install the bot →
Try it Bot How it works Pricing Contact GitHub ↗ Open App →
Deterministic. 96 of 96 blocked graphs crashed PyTorch

Claude generates models.
Nobody verifies models.

Neurarch is the agent in your repo that plans every model change, blocks the ones that will not run, and remembers what trained. One command gives you a plan you can share.

The one line in that plan no frontier model can write, because the data is not in it: history Last time this structure trained here: valAcc 0.9865, 3 epochs, <$0.01, 34s, 2026-09-04 That row is real, off a pull request, out of your own organisation's ledger and nobody else's.

For researchers reading papers teams shipping models without an ML research org behind them teams onboarding new hires
Install the bot → Paste a model, get the plan
Free, no signup to start building No credit card required You own the exported code API keys never leave your browser

Five ways in

Same verifier behind all five. One is a recording of the whole loop; the other four are places to run it where you already work.

Watch a real run One recorded run end to end: designs proposed, the broken ones refused, one paid GPU run, and the trained model in your browser. www.neurarch.com/watch Try it Paste a model, get the card. Nothing to install, no account. www.neurarch.com/try The bot Two files in your repo, and every pull request gets a plan. uses: neurarch-ai/neurarch-bot@v0 MCP server Your coding agent gets the graph, and the checks, as tools. claude mcp add neurarch -- npx -y neurarch-mcp /abs/path/to/model.py The API plan, check, rank, fix, train, history. Over HTTP, no model call. curl -sS https://www.neurarch.com/api/v1/check

One command traces a forward pass and prints a plan: params, will it run, GPU fit, cost. Real output, real link.

zsh
$ neurarch-trace torchvision.models:resnet18 --input 1,3,224,224 --plan --share
Plan: resnet18   68 layers · 11.7M params · fingerprint 9ea7d4ae
  shape     [3,224,224] -> [1000]
  will run  yes (0 blockers, 0 warnings)
  fits      T4 16GB yes · A100 40GB yes · H100 80GB yes
  train     ~$0.11 / 6m on A10G (24GB)   (10 epochs, batch 32, 50k samples assumed)
Share: https://www.neurarch.com/p/r7rq5c4jt2
pip install neurarch-trace
182 layer types
75 macro blocks
36 model templates
41 structural checks
The problem

Building the model is easy.
Knowing it's the right one is hard.

The wrong design passes

It compiles, forward-passes, and trains to the last epoch. No stack trace, no failing check, nothing to grep for. You learn it was the wrong one after you paid for it.

Training is the only oracle

To tell two legal designs apart, you train both. That is the price of one answer to one question, and you ask it again every time you change something.

No static score ranks them, ours included

Across 24 designs we trained on a real GPU, outcomes ran from 6.8% to 97.2% while static scores barely moved: Spearman 0.17, essentially uninformative. We publish our own number because closing that gap is the product. See the 24 rows →

Design-time lint

Watch a bug a type checker can't see die in the editor

Head divisibility is one of 41 structural checks that run as you build. See real bugs caught before the GPU bill →

neurarch.com

Real session, no cuts: a model on the canvas, then which GPU it fits, in 17 seconds.

See it in action

One continuous take on the live app, captioned: sketch layers by hand while Tab autocompletes and wires the next one, let the AI complete the design without touching your layers, or describe the task and the wizard loads BERT. Read the design-time serving physics, KV cache, hardware fit and cost per million tokens before a single training step. Run the Model Audit with its GPU-fit table and GQA / MLA what-ifs, flip on the roofline overlay, let the agent deepen the encoder, then training, evaluation, deploy, and the same graph as PyTorch and TensorFlow side by side. Browse the Studio's derived artifacts and the community gallery, audit any HuggingFace model (Qwen-2.5 7B), open it on the canvas folded to blocks, and export a publication-ready figure. Design, lint and cost it before you train.

A real GPU run costs cents, not a mystery bill

T4 · $0.59/h · billed by the minute this run: ~3 min · $0.04 98.0% val accuracy · 15 epochs

Recorded against production on a free-to-start account: the task wizard assembles a CNN, the pre-launch card quotes the run ($0.04 on a T4, auto-stop at your budget), real loss/accuracy curves stream back from the GPU, and the pipeline carries the run through Evaluate (which flags the model's own overfit) into the Deploy advisor's pricing ($0.16 per million requests). Only edit: the mid-run wait is fast-forwarded 16x. No card required for the free path: the same project exports as a Colab notebook that trains on a free T4 and reports back, and simulation mode stays free for design iteration.

A 30-second walkthrough

What it actually feels like

YOU
Build a vision transformer for CIFAR-10
NEURARCH
ViT-Tiny, patch embedding → 12× attention block → CLS head. ~5.5M params, all shapes verified. I also caught a Softmax feeding into CrossEntropyLoss (which already applies one) and removed it. → canvas updated · 1 structural bug auto-fixed · 0 lint errors · ready to train
YOU
Make it deeper, drop the CLS token, use GAP head
NEURARCH
Diff applied: 12 → 24 attention blocks, removed CLS, swapped to global average pooling. → params 5.5M → 11.0M · gradient flow OK · advisor: consider stochastic depth above 18 layers
YOU
Train it on a free GPU, then export PyTorch
NEURARCH
Notebook running on a free Colab T4, epoch 30/50, val_acc 78.4%, curves streaming back. Diagnosing now: slight underfitting, suggesting +RandAugment. Generated vit_cifar.py + training script. → download · or one-click re-run with the fix

Every step above is a real action in the app, not a script. Try it yourself →

What people say

From early beta users

Names anonymized while we collect attribution permission · roles representative of actual users

★★★★★
"The biggest defense of this product is workflow integration, drop an arXiv link or a PyTorch file and land in an interactive ML-specific editor. The long-term moat is the deep engineering: AST parsers, intelligent graph routing, academic export pipelines."
JM
J.M.
Member of Technical Staff · frontier AI research lab
★★★★★
"This aligns directly with what we're building. It would save us most of the early-stage architecture-design work, the part our team currently rebuilds from scratch every time on Transformer-based architectures."
LK
L.K.
CEO · early-stage AI startup
★★★★★
"The idea and the visual interface are great, clean and intuitive. I uploaded a research paper, and they shipped 7 fixes to the import pipeline within 24 hours. If the full paper-to-runnable-code path lands end-to-end, this tool is unbeatable."
TW
T.W.
CS PhD researcher · US research university
Use cases

What are people building?

Industrial defect detection

Catch defects straight off the camera feed. Small enough to run on the line's edge hardware.

"Build a CNN classifier for defect detection in manufacturing images"
Try this prompt →

Support ticket routing

Read every ticket. Route it by urgency and department. Walk away with the fine-tuning code.

"Route customer support tickets by urgency and department"
Try this prompt →

IoT anomaly detection

Watch the sensor stream. Flag the failure before it happens. Variable-length sequences handled.

"Anomaly detector for IoT sensor data"
Try this prompt →

Replicate a paper

Paste the arXiv link. Get the actual wiring, not Figure 2. Diff it against your baseline and train.

"Replicate ResNet-50 from the original paper"
Try this prompt →

E-commerce recommendations

Two towers, user and item embeddings. Trained on your interactions. Shipped as code you own.

"Recommendation model with user and item embeddings"
Try this prompt →

On-device ML

Fits in 10 MB. Runs on a phone CPU. Quantization preview and hardware-fit analysis built in.

"Compact transformer for on-device NLP on mobile CPU"
Try this prompt →
How it works

Four steps from design to a shipped model

01

Design the architecture

Sketch it yourself, or let the agent draft one from a spec like "10-class image classifier on 224×224 RGB". Either way the graph is typed and yours to edit.

02

Lint it before you train

41 structural rules and shape propagation run as you build: head divisibility, shape mismatches, dead paths, blown param and KV budgets. The wiring bug shows up in the editor, not three hours into a GPU run.

03

Train & get AI diagnosis

Run training. AI reads the loss curves, diagnoses overfitting or underfitting, and applies targeted fixes.

04

Export & ship

PyTorch, Keras, ONNX, Jupyter notebooks. Clean readable code that runs without any Neurarch import.

Integrations

One graph in the middle.
Generated glue everywhere else.

Everything on the left lands in the same typed graph. Everything on the right is generated from it: runnable, readable, yours. No third-party tool to host, no plugin to babysit.

Bring
arXiv paper URL
HuggingFace model ID
PyTorch source
.onnx / .safetensors
Paper figure screenshot
Plain-English spec

The typed graph

shapes · lint
params · serving cost

Ship
PyTorch + Keras code
Training notebook
Free GPU run (Colab / Kaggle)
Docker + FastAPI bundle
ONNX + browser inference
TikZ figure + BibTeX
Features

What's in the box

The basics

Typed graph with shape propagation

182 layer types + 75 macro blocks. Tensor shapes propagate automatically. Dimension mismatches caught before you train.

AI agent that sees your canvas

The agent has access to your selected layers, the shape trace, and the lint output. So when you ask "why's this exploding?", it answers from your model, not a textbook one.

Architecture advisor (35 rules)

Catches vanishing gradients, missing normalization, GQA head mismatches, MoE aux-loss, overfitting risk, ordering errors before you waste compute. See the full rule book →

arXiv paper → canvas

Paste a paper URL. The agent parses the architecture and builds it on the canvas. Diff against your version.

HuggingFace import

Paste an HF model ID, get a clickable architecture. Useful when you want to read what BERT or LLaMA actually does instead of skimming the config.

Production code export

PyTorch, Keras, ONNX, Jupyter notebooks, PDF reports. Full training script with optimizer and scheduler.

Real GPU training, free or managed

Drop a CSV, pick a HuggingFace dataset, or upload an image-zip. Train free on a Colab / Kaggle GPU via a one-click notebook, or run in-app on managed GPUs (T4 to H200) with live curves. AI diagnoses the results and proposes the patch.

Snapshots + run history

Save architecture checkpoints. Diff any two versions. Compare training runs side-by-side, mark a best run, re-use a config in one click.

Real-time collaboration

Live cursors, shared canvas, team model library. Built on our own WebSocket layer, no Pusher / Liveblocks dependency. Launching with Pro Plus.

The moat deep engineering

Academic export pipeline

Every layer carries provenance. One canvas state generates a TikZ figure, a "Methods" paragraph, and a BibTeX file with the right entries cited. Authoring path for ML papers, not toy demos.

Paper figure → architecture

Drop a screenshot of any architecture diagram. Vision LLM rebuilds nodes + edges (residuals included). Skip-edges get bowed routing so the result reads like the original figure.

Cross-validation (Paper ↔ Code)

Point at a paper and your nn.Module source. Per-layer drift report: dimensions, activations, ordering, missing residuals. Catches re-implementation bugs before review.

Canvas ↔ Code identity merge

Edit code in Monaco, parse back, the canvas only updates what changed (matched by name first, type-ordinal second). BERT's two norm layers don't collide; residuals survive when you parse the edited code back.

Deploy advisor

Latency / FLOPs / size / energy on 7 platforms (mobile, browser-WebGPU, Coral edge-TPU, A100, …). Quantization variants table, perf-budget gauge, smell detector with clickable layers, comparison vs LeNet → Llama-3.

Browser ONNX inference

Export a real .onnx straight from the canvas, drop it back in the browser to run inference live (WASM + WebGPU). Top-5 softmax bar, no server round-trip. CSV upload auto-infers the task before you even build.

Why Neurarch

AutoML hands you a model.
Neurarch hands you the check that says it's wrong.

Capability Neurarch HF AutoTrain Google AutoML Write code
Typed graph you can edit (not a read-only view)
Semantic checks (head divisibility, norm placement, aux loss)
Serving physics before you train (KV cache, GPU fit)
Architecture lint (gradient / overfit / order)
Natural language → architecture
Import arXiv paper → canvas
Import any HuggingFace modelmanual
Import .onnx / .safetensors → editable canvas
PyTorch + ONNX code export (you own it)
AI reads training curves + targeted fixeslimited
Real-time team collaborationPro+ soon
Free to starttrial

AutoML hands you a black box and a number. Great if the number is good enough. If it isn't, you have nowhere to start debugging. Neurarch shows you the layers, the shapes, and the code, so when it goes wrong, you know which line to change.

For when you'd reach for a sketchpad or print(model)

Capability Whiteboard / Excalidraw print(model) Netron Neurarch
See full layer graphmanualtree only
Tensor shapes verified end-to-endif forward runs
Edit graph + re-export coderead-only
Architecture lint (gradient/overfit/order)
Build from prompt or arXiv paper
AI reads training curves + applies fixes
Shareable URL of your architecturescreenshotfile only

For when you'd ask ChatGPT or Copilot to just write the PyTorch

Dimension LLM writes PyTorch Neurarch
What the model lives asfree-form codetyped graph + the same code
Edits checked before they apply41 structural checks
Head-dim / GQA / merge bugs surfaceat GPU timein the editor
Cost to verify one edita forward pass0.27 ms, one laptop core
Does the verdict track reality?untested96/96 blocked graphs crash PyTorch
Do passing designs actually train?find out on the GPU24/24 trained end-to-end on a managed T4
Blast radius of a changegrep and hopeimpact analysis per edit
You own the code

An LLM writing PyTorch is fast and often right. The problem is the edit nobody checks: the head count that stops dividing the embedding dim, the Linear that no longer matches its upstream width. Those compile, pass review, and fail on the GPU you already paid for. Because Neurarch keeps the model as a typed graph, every edit (yours or the agent's) is verified in under a millisecond before it lands. In a 264-graph study, all 96 designs the verifier blocked crashed in PyTorch, and all 80 it passed ran clean. The loop now closes upward too: in the grounded benchmark split, all 24 verifier-passed designs (18 by frontier models, 6 human references) trained end-to-end on a managed T4 through the production trainer. On 4 of 5 classification tasks, the best verifier-passed model-generated design finished at or above the human reference's accuracy after identical real training; on the other task, every generated design trained to noise. The designs that trained to noise are the failure class no static check can see, which is exactly why training outcomes feed back into the checks. The 41 checks are public, rule by rule.

Already using Netron to inspect models? Drop any .onnx or .safetensors file onto Neurarch, we reconstruct the full graph as editable canvas nodes. Modify layers, fix architecture issues, run training, and export clean PyTorch code. Netron shows you what a model is. Neurarch lets you change it.
Already browsing HuggingFace to find a model? HuggingFace stores and displays models that already exist, read-only. Neurarch sits one step upstream: design a new architecture, catch dimension / param / cycle bugs before you burn a GPU run, and export runnable training and deploy code. Then push to HuggingFace to host. HuggingFace shows you a model that exists. Neurarch helps you design one that doesn't, yet.
Pricing

Free to start. Pay when you scale.

Beta pricing, rates lock in for life when you subscribe during early access.

Free
$0 forever
Full canvas, AI agent, and core code export. No card, no time limit.
  • Full graph editor with 182 layer types + 75 blocks
  • 15 hosted AI agent messages per day (Qwen3-Next 80B)
  • Claude via your own Anthropic key
  • PyTorch, Keras, ONNX, training script export
  • Architecture advisor (35 rules)
  • 5 local snapshots + URL sharing
Start for free →
Most popular
Pro
$19 /mo
Everything single-user, every importer, every export, every research tool.
  • 500 hosted AI agent messages / month (DeepSeek V3)
  • Imports: HuggingFace, arXiv paper, AI code parser
  • Exports: Notebook, Report, Slides, Study Guide, Model Card
  • Cost estimator, quant preview, receptive field, scaling calc
  • Architecture compare + profiler import + fusion hints
  • 20 snapshots + cloud save
  • 7-day free trial
Start free trial →
Pro Plus
$39 /mo
Pro + Claude quota, the heavy research suite, and team collaboration as it lands.
  • 2,000 hosted AI agent (DeepSeek V3) + 200 Claude Sonnet / month (no key needed)
  • Real-time team collaboration + workspaces (soon)
  • Unlimited snapshots
  • Loss landscape, logit lens, attribution explorer
  • Live training dashboard, training replay, Bayesian opt
  • GPU profiling, auto-viz generator, custom layer editor
  • Architecture marketplace + browser
  • SSO + audit log + priority support
  • Everything in Pro · 7-day free trial
Start free trial →

Need invoicing, a custom seat count, or on-prem? contact us →

Privacy & data

Your code never leaves your browser

No accounts required for the canvas, agent, or export. When you BYOK, the request goes browser → provider, we're not in the path.

01

Type or paste in the canvas

Architecture, prompts, and pasted code live in browser memory only. Refresh the tab to wipe them.

02

API key stays in your browser

Your Anthropic / Gemini key is stored in your browser's localStorage, sent directly to the provider, never to our servers, and wiped when you sign out or switch accounts.

03

BYOK calls go direct to the provider

Browser → Anthropic / Google. We're not a man-in-the-middle. Open the network tab and verify it yourself.

04

Nothing persisted unless you opt in

Local snapshots stay in browser memory. Cloud save (Pro) and team workspaces (Pro Plus) are explicit actions, one-click delete from the dashboard at any time.

From the lab

What we're measuring in public

Calibration Aug 2026

How badly a passing score predicts a trained model

24 designs that all passed the verifier and all trained end to end on one T4 finished between 6.8% and 97.2%. Spearman between static score and trained outcome: 0.17. Every row is on the page, both ways of pooling it, and the part we have not closed yet.

Read the calibration →
Arena Jul 2026

Can frontier models design valid neural networks?

The Arch-Bench arena shows a model a design spec and a starting graph, takes structured edits, and grades the result with a deterministic verifier. No human judge, no LLM judge, no GPU. Every leaderboard row reproduces with one command.

Enter the Arch-Bench arena →
Environment Jul 2026

Train an agent against the verifier

A gym for architecture design: reset to a design spec and a starting graph, take typed structured edits, get back a deterministic, verifiable reward. No LLM judge, no GPU. GRPO-ready, one command to run.

Open the environment →
API Aug 2026

The verdict, over HTTP

POST a graph to /api/v1/check and get back what the app shows a human: can it train, what would the run cost, where should it serve, and what is still a person's decision. Same code path, no browser, no LLM. Built for an agent that generates a design and needs to know if it holds.

Read the API docs →
Field notes Jul 2026

Bugs caught before the GPU bill

Structural bugs don't fail in code review. They fail at forward time, on hardware you already paid for. A running list of real bugs the structural checks intercepted at design time.

Read the case list →
Calculator Jul 2026

The serving math, out in the open

KV cache per token, GPU fit, cost per million tokens. The same estimators the in-app Model Audit runs, validated against published serving numbers.

Open the calculator →
FAQ

Common questions

Both. By default it runs an architecture-aware training simulation, curves reflect your architecture's real quality signals (dropout, residuals, normalization, depth) so the AI's interpretation is meaningful. For real GPU training there are two paths: the free one-click Colab / Kaggle notebook (trains on a free GPU, streams curves and checkpoint back to your account), or in-app managed GPU runs on T4 through H200 with live curves and per-plan budget caps. Either way the exported PyTorch code runs anywhere.
No. The "Problem → Model" wizard asks a few questions and the agent picks layers for you. But you'll get more out of it once you start recognising what each block does. The canvas and the advisor are designed to teach you that as you go, not to hide it from you.
No. API keys live in browser session storage and never reach our servers, BYOK calls go straight from your browser to the provider. We proxy the hosted AI agent (Qwen3 on Free, DeepSeek V3 on paid plans, via OpenRouter) for signed-in users so we can enforce per-plan quotas (15 messages/day on Free, 500/mo on Pro, 2,000/mo on Pro Plus). Claude proxy is Pro Plus only, Free and Pro users supply their own Anthropic key.
ChatGPT gives you a code block you can't see inside. Neurarch gives you a typed graph where every layer is inspectable, shapes propagate across the full model, and 41 structural rules run automatically. You can import from HuggingFace or arXiv, run training, get AI diagnosis of the curves, and export. The agent is architecture-aware, not just code-aware.
Yes. Use the code importer to load an existing architecture, or load a HuggingFace model ID to visualize it. The exporter generates clean nn.Module classes you can paste into any existing codebase.
Yes. Cancel from the billing portal anytime. Pro access continues until the end of the period. Full refund within 7 days of first payment if you're not satisfied.
Roadmap

What's built & what's coming

Typed-graph editor with 182 layer types + 75 blocks Shipped

Authoring surface with automatic tensor shape propagation, 41 structural checks, and drag-and-drop editing.

AI architecture agent Shipped

Tell the agent what you're building. It picks the layers, wires them up, and patches its own mistakes when the advisor flags something.

arXiv + HuggingFace import Shipped

Paste any arXiv URL or HF model ID and the architecture appears on your canvas instantly.

PyTorch / Keras / ONNX / Notebook export Shipped

Production-ready code, full training scripts, Jupyter notebooks, you own the output.

Real GPU training, free Colab / Kaggle loop Shipped

One-click notebook that trains on a free Colab or Kaggle GPU and streams loss curves, eval metrics, and the best checkpoint back to your account. Drop a CSV, pick a HuggingFace dataset, or upload an image-zip. AI diagnosis of the results.

Managed GPU training (Modal-backed) Shipped

The same runs in-app on managed GPUs, T4 up to H200, with live loss curves and no notebook round-trip. Per-plan runtime and budget caps enforced server-side.

Real ONNX export + browser inference Shipped

Export the canvas to a real .onnx file, right in the browser. Drop any .onnx back in to run inference live (WASM + WebGPU). Plus Docker / FastAPI / training-script bundles for deploy anywhere.

Academic export pipeline Shipped

Every layer carries provenance (paper / HF / manual). Generate a TikZ figure, a "Methods" paragraph, and a BibTeX file with the right entries cited, straight from canvas state.

Team collaboration In progress

Live cursors, shared canvas, team model library. Pro Plus feature launching with Stripe billing.

LoRA / PEFT fine-tuning on LLMs Shipped

Pick a base model, pick a HuggingFace instruction dataset, and the managed GPU runs the LoRA, QLoRA or DoRA loop with the loss curve streaming back live. The trained adapter is kept, so you can serve it behind an OpenAI-compatible endpoint in one click. Open-weight bases (Qwen, Phi, Mistral) work as they are; gated ones need your own HuggingFace token.

Q4'26

One-click push to HuggingFace Hub Coming Q4 2026

Trained model + auto-generated card → pushed to your HF account in one button. Local Docker / FastAPI deploy bundles already work today.

Start free today

Full canvas + AI agent + code export on the free plan.
Drop your email to get Pro updates and early access pricing.

No spam. Or just open the app now →

Share your model

Add this badge to your README, notebook, or paper to show the architecture was built with Neurarch.

Designed with Neurarch
[![Designed with Neurarch](https://img.shields.io/badge/Designed%20with-Neurarch-5b6cff?style=flat&logo=data:image/svg%2bxml;base64,PHN2ZyB3aWR0aD0iMTYiIGhlaWdodD0iMTYiIHZpZXdCb3g9IjAgMCAxNiAxNiIgZmlsbD0ibm9uZSIgeG1sbnM9Imh0dHA6Ly93d3cudzMub3JnLzIwMDAvc3ZnIj48Y2lyY2xlIGN4PSI4IiBjeT0iOCIgcj0iNyIgc3Ryb2tlPSJ3aGl0ZSIgc3Ryb2tlLXdpZHRoPSIxLjUiLz48cGF0aCBkPSJNNSA4aDZNOCA1djYiIHN0cm9rZT0id2hpdGUiIHN0cm9rZS13aWR0aD0iMS41IiBzdHJva2UtbGluZWNhcD0icm91bmQiLz48L3N2Zz4=)](https://neurarch.com)

Click to copy Markdown

Contact

Talk to us

Bug reports, feature requests, partnerships, design partners, investor intros, fastest way to reach us is email or GitHub.

Email us
neurarch.ai@gmail.com
Book a demo
15-min walkthrough
GitHub
github.com/neurarch-ai
Draft feedback
Opens your mail client