PhonePredict — Mobile Price Predictor

A polished, two-part app:

• Frontend — Vite + React + React Router + Tailwind CSS + GSAP
• Backend — FastAPI + scikit-learn (RandomForest) + Tavily web search + Claude for structuring results

Give it a company (optional), rating, RAM (GB) and storage (GB) → it predicts a fair market price, then finds and structures 5 real phones worth comparing it to.

---

1. Backend setup

cd backend
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

# Train the model (generates backend/app/data/model.pkl)
python train_model.py

# Copy env template and (optionally) add API keys
cp .env.example .env

# Run the API
uvicorn app.main:app --reload --port 8000

The API is now live at http://127.0.0.1:8000. Docs at http://127.0.0.1:8000/docs.

Optional API keys (.env)

The app works with zero keys — it falls back to a curated local phone catalogue for recommendations (recommendation_source: "local_fallback" in the response). Add these to unlock live results:

Variable	Purpose	Get it at
TAVILY_API_KEY	Live web search for current phones near the predicted price	https://tavily.com
ANTHROPIC_API_KEY	Structures raw search results into clean JSON recommendations	https://console.anthropic.com

With only TAVILY_API_KEY set, you still get results (recommendation_source: "tavily"), just less neatly structured. With both set, you get "tavily+llm" — the full pipeline.

2. Frontend setup

cd frontend
npm install
cp .env.example .env   # points VITE_API_BASE_URL at your backend
npm run dev

Open http://localhost:5173.

3. How a request flows

1. User submits company / rating / RAM / ROM.
2. POST /api/predict → the trained RandomForest model predicts a price.
3. The predicted price + specs are used to query Tavily for current phones in that range.
4. The raw search results are passed to Claude with a strict JSON-only prompt, which returns exactly 5 structured recommendations (name, company, price, specs, reason, source_url).
5. If Tavily or the LLM step isn't configured/fails, the API transparently falls back to ranking a local curated phone dataset by spec/price similarity — the UI always gets a complete result.

4. Retraining on real data

train_model.py currently trains on a synthetic-but-realistic pricing formula (see COMPANY_PREMIUM, price_formula). To use real data, replace generate_synthetic_data() with pd.read_csv("your_data.csv"), keeping columns Company, Rating, RamSize_GB, RomSize_GB, Price.

Project structure

backend/
  app/
    main.py            FastAPI app + /api/predict route
    model.py            Loads model.pkl, runs inference
    schemas.py           Pydantic request/response models
    tavily_client.py     Live web search
    llm_structurer.py    Claude-based result structuring
    recommend.py         Local fallback recommender
    data/
      phones.json         Curated sample phone catalogue
      model.pkl           Trained model (generated)
      companies.json      Companies model was trained on (generated)
  train_model.py
  requirements.txt
  .env.example

frontend/
  src/
    components/   PriceDial (signature gauge), PredictForm, PhoneList, Navbar, ThemeToggle
    pages/        Predictor, History, About
    theme/        Light/dark ThemeContext
    lib/          api.js, history.js (localStorage)