Skip to content

project-vision.md

Latest commit

 

History

History
153 lines (112 loc) · 7.65 KB

project-vision.md

File metadata and controls

153 lines (112 loc) · 7.65 KB

AI Model Advisor - Project Vision

Mission

Find AI models you can run in the browser, on mobile, or at the edge — no server required. Focused on sustainability, privacy, and simplicity.

Core Principles

  • Edge First: Recommend models that run on the user's device — no data center needed
  • Privacy First: Runs locally in browser, no external data transmission
  • Environmental Consciousness: "Smaller is better" — on-device inference eliminates server-side compute entirely
  • Simplicity First: Prove concept before adding complexity
  • Open Source: Community-driven development

Current Status

✅ MVP Complete (Phase 1)

  • ✅ SvelteKit project structure with static site generation
  • ✅ Browser-based embedding classification (MiniLM, 98.3% accuracy)
  • ✅ Tiered recommendation engine (Lightweight/Standard/Advanced)
  • ✅ Environmental impact scoring (1-3 scale)
  • ✅ Model accuracy filtering (50-95% threshold)
  • ✅ Lightweight model (~23MB) enables mobile support
  • ✅ Automated model dataset updates (daily via GitHub Actions)
  • ✅ PWA with offline capability (installable on desktop & mobile)
  • ✅ Full accessibility support
  • ✅ 152 curated models across 4 categories, 10 subcategories

Live: https://ismaelmartinez.github.io/ai-model-advisor

Roadmap

Phase 1: MVP ✅ Complete

Goals: PWA with tiered recommendations and environmental scoring

Achieved:

  • <1 second to recommendations (target was <30s)
  • 4 AI task categories, 10 subcategories, 98.3% classification accuracy
  • ~40KB bundle size (target was <2MB)
  • Full accessibility support
  • Mobile & desktop PWA support
  • Daily automated model updates from Hugging Face

Phase 2: Refocus on Edge/Browser AI ✅ Complete

Goal: Sharpen the project's identity around edge-deployable models. Reposition from "general AI model advisor" to "find AI you can run without a server."

Why now: The edge AI ecosystem has matured — Transformers.js v4 supports ~200 architectures, WebGPU is broadly available, and small models (1-4B) are closing the quality gap. No tool answers: "I want to do [task X] without a cloud API — what are my options?" See edge-browser-ai-direction.md for full analysis.

Phase 2 is organized into three parallel workstreams plus one independent optimization:

Workstream A: Data & Metadata Enhancement ✅

Can run independently — no UI or build changes needed

  1. Tag models with verified edge runtime support — Add runtime metadata (framework, backend, tested/untested) to the top 40-50 edge-compatible models already in the database
  2. Add task maturity labels — Classify each subcategory as "excellent", "good", "limited", or "emerging" for edge deployment (see research doc Appendix tables)
  3. Curate for deployability — Surface edge-compatible models first; de-emphasize models requiring dedicated GPU servers (keep them accessible, just deprioritized)
  4. Expand edge model coverage — Add verified browser-runnable models from Transformers.js, TensorFlow.js, MediaPipe, and WebLLM ecosystems

Workstream B: UX & Messaging Refresh ✅

Can run in parallel with Workstream A

  1. Update positioning and copy — Hero text, descriptions, and environmental messaging to reflect edge/browser focus
  2. "Edge-Ready" badge — Visual indicator on models with verified browser/mobile support
  3. Structural environmental story — Frame on-device AI as an architecture choice: no data center, no network round-trip, amortized hardware, privacy bonus
  4. Honest task maturity display — Show users which tasks are well-supported at the edge vs. experimental

Workstream C: Build-Time Optimization ✅

Fully independent — can run in parallel with A and B

  1. Pre-compute reference embeddings — Distill the ~320 reference example embeddings at build time into a static JSON file, so the browser skips embedding them on first load. Cuts initial classification latency significantly.

Workstream D: Cross-Browser & Testing ✅

Can start in parallel, but benefits from A completing first

  1. Cross-browser testing — Validate Firefox, Safari, Edge support
  2. Fix any edge cases — Address test failures found during broader browser testing
  3. Improve test coverage — Add tests for new runtime metadata and edge-ready flags

Parallelism map:

         ┌─── A: Data & Metadata ───────────────┐
         │                                       │
Start ───┼─── B: UX & Messaging ────────────────┼──→ Phase 2 Complete
         │                                       │
         ├─── C: Build-Time Embeddings ──────────┤
         │                                       │
         └─── D: Cross-Browser (start early, ────┘
                  benefits from A finishing)

Phase 3: Make It Actionable

Goal: Turn recommendations into starting points for implementation.

Depends on Phase 2 data work being complete.

  1. Code snippets — "Get started" snippets for Transformers.js, ONNX Runtime Web, TensorFlow.js for each recommended model
  2. Runtime comparison — Side-by-side framework comparison showing which runtimes support each model
  3. Performance estimates — Cold start time, inference latency, memory footprint from community benchmarks and our own testing
  4. "Can I Run This?" feature — Input a model name or Hugging Face URL → get an edge deployment feasibility report (estimated browser memory, compatible runtimes, quantization options, WebGPU vs WASM expectations)

Phase 4: Community & Ecosystem

Goal: Build the dataset that makes this tool indispensable.

  1. Community-reported benchmarks — Users submit real-world performance data for models they've tested
  2. Automated compatibility testing — CI job that tests top models in headless Chrome to verify browser runnability
  3. Framework update tracking — Monitor Transformers.js, ONNX Runtime Web releases for new model support
  4. Expanded model coverage — Systematically test and add models from HF Hub that have ONNX/WASM exports

Technical Architecture

User Input → Embedding Classification (MiniLM) → Model Selection → Environmental Scoring → Recommendations

Fallback Chain: Embedding Similarity → Semantic Matching → Keyword Matching → Default Category

Success Metrics

  • Speed: <1 second query to recommendation ✅
  • Accuracy: 98.3% classification accuracy ✅
  • Performance: ~40KB bundle, <1s load time ✅
  • Mobile Ready: ~23MB model, PWA installable ✅
  • Edge models with runtime data: 25+ models with verified runtime metadata ✅
  • Code snippet coverage: 0% → 80% of recommended models (Phase 3 target)
  • Adoption: Target 2,000+ monthly users within 6 months of Phase 2 launch

Contributing

  1. Test with real use cases
  2. Curate model data (especially edge runtime metadata)
  3. Improve algorithms
  4. Report browser compatibility findings
  5. Promote sustainability focus

See CONTRIBUTING.md for details.

Documentation

Impact Goals

Help developers find and deploy AI models that run on their users' devices — eliminating server-side compute, protecting user privacy, and reducing the environmental footprint of AI inference.
⚡