Organization card

README.md

@@ -1,10 +1,67 @@
 ---
 title: README
-emoji: 🏆
+emoji: 🔵
 colorFrom: blue
-colorTo: purple
+colorTo: indigo
 sdk: static
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<p align="center">
+  <img src="https://huggingface.co/spaces/nodd-repo/README/resolve/main/logo.svg" alt="nodd" width="96">
+</p>
+
+<h1 align="center">nodd</h1>
+
+<p align="center">
+  <b>Tiny, calibrated classifiers that run in your browser.</b><br>
+  One decision → one ~24 MB model: offline, no per-call cost, and it knows when it is unsure.
+</p>
+
+<p align="center">
+  <a href="https://michaljach.github.io/nodd/">Live demo</a> ·
+  <a href="https://michaljach.github.io/nodd/docs.html">Docs</a> ·
+  <a href="https://github.com/michaljach/nodd">GitHub</a> ·
+  <a href="https://www.npmjs.com/package/@nodd/browser">@nodd/browser</a> ·
+  <a href="https://www.npmjs.com/package/@nodd/node">@nodd/node</a>
+</p>
+
+---
+
+Each model here is a MiniLM encoder fine-tuned for **one** task, exported to int8 ONNX and
+checked for parity against Python. Every prediction comes with a **calibrated** confidence: above
+the threshold the model decides alone, below it you escalate to a bigger model.
+
+| Model | Decides | Labels | Test macro-F1 | Handled alone* |
+|---|---|---|---|---|
+| [comment-moderation](https://huggingface.co/nodd-repo/comment-moderation) | Can this blog comment be published? | `ok` `spam` `toxic` | 0.952 | 94% |
+| [prompt-injection](https://huggingface.co/nodd-repo/prompt-injection) | Is this message trying to hijack an AI assistant? | `safe` `injection` | 0.961 | 100% |
+| [sentiment](https://huggingface.co/nodd-repo/sentiment) | What feeling does the writer express? | `positive` `neutral` `negative` | 0.882 | 90% |
+| [support-triage](https://huggingface.co/nodd-repo/support-triage) | Which queue should this support message go to? | `billing` `bug` `account` `how_to` `feature_request` | 0.901 | 44% |
+
+<sub>*Share of validation inputs above the confidence threshold, which is set for ≥ 97% precision.
+Trained on synthetic data: expect lower numbers on real traffic.</sub>
+
+### Use
+
+```sh
+hf download nodd-repo/comment-moderation --local-dir public/models/comment_moderation/v3
+npm install @nodd/browser   # or @nodd/node on a server
+```
+
+```ts
+import { nodd } from "@nodd/browser";
+
+const m = await nodd.load("/models/comment_moderation/v3");
+const d = await m.decide("Buy cheap followers at ...");  // { label: "spam", confidence: 0.99, ... }
+if (!m.isConfident(d)) { /* escalate to a bigger model */ }
+```
+
+### Make your own
+
+Write a short YAML spec (the input and a fixed set of labels), and `nodd` collects data, labels it
+with a teacher model, trains, calibrates and exports a browser-ready model, all on CPU.
+
+```sh
+uv run nodd run examples/comment_moderation.yaml
+```