Client-side AI Experiments

Client-side AI playground — browser-native ML demos powered by WebAssembly. No servers, no API keys, no data leaves your device.

Demos

Demo	Description	Tech
Vector Search	Sub-ms nearest-neighbor search over 1,000 vectors using HNSW graph traversal in a 42 KB WASM microkernel	@ruvector/rvf-wasm
Emoji Finder	Semantic emoji search — type a feeling or concept and find matching emojis by meaning, not keywords	Feature hashing + WASM HNSW
Semantic Search	Semantic movie search in Danish — type a description and find matching films by meaning using in-browser embeddings	Transformers.js (gte-small)
Dansk NER	Find persons, organizations and locations in Danish text using a ModernBERT model (~144 MB) running entirely in-browser	Transformers.js + ONNX Runtime
Rock Paper Scissors	Play RPS against an adaptive AI that learns your patterns in real-time using ruvnet's SONA self-optimizing neural architecture in WASM	SONA + WASM

Quick start

npm install
npm run dev       # Vite dev server on http://localhost:5173
npm test          # Unit tests (vitest)
npm run build     # TypeScript lib → dist/

Project structure

src/              TypeScript library (WASM init, loaders, HNSW engine, metrics)
examples/         Browser demos (vanilla JS, each self-contained)
tests/            Unit tests + Playwright e2e tests
config/           Vite, Vitest, Playwright, TypeScript configs
plans/            Session notes and plans

Future ideas

Things to try next — all running client-side in the browser:

Image processing client-side

Ruvector CNN samples: https://github.com/ruvnet/RuVector/tree/main/docs/cnn

WASM and working. Extract other simple usecases.

NLP & text

• Sentiment analysis — classify Danish/English text as positive/negative/neutral using a small ONNX model
• Text summarization — extractive summarization of articles using sentence embeddings + MMR selection
• Spam / toxicity filter — lightweight classifier for comment moderation, no server round-trip
• Keyword extraction — TF-IDF or RAKE running in WASM for instant keyword highlighting
• Language detection — identify language from a text snippet using character n-gram vectors

Search & retrieval

• Document search — embed paragraphs with a small model (MiniLM), index with RVF HNSW, search by meaning
• FAQ matcher — semantic match user questions to a knowledge base of answers
• Image search by caption — CLIP-style text→image retrieval using quantized ONNX vision encoder
• Code search — embed code snippets and search by natural language description

Classification & tagging

• Image classification — MobileNet or EfficientNet-Lite in ONNX for real-time object recognition from webcam
• Document categorization — auto-tag uploaded documents by topic using zero-shot classification
• Auto-tagging for photos — detect objects/scenes in browser using small vision models

Generative / interactive

• Text completion — small GPT-2 or Phi-3-mini running in ONNX for local text generation
• Translation — Danish↔English using Helsinki-NLP OPUS models via Transformers.js
• Speech-to-text — Whisper tiny/base in ONNX for in-browser transcription from microphone
• Text-to-speech — client-side TTS using VITS or similar small vocoder models

Tooling & infrastructure

• WASM vector DB benchmark — compare RVF HNSW against brute-force and other ANN approaches at various scales
• Embedding cache — persist computed embeddings in IndexedDB for instant reload
• Model size explorer — interactive comparison of model size vs. accuracy tradeoffs for browser deployment
• Progressive model loading — stream large models in chunks, become interactive before full download
• WebGPU acceleration — use WebGPU for matrix multiplication when available, WASM fallback otherwise