Türkische Fahrzeugbewertungsplattform Builder

## Step 1 – Product & System Design ### High-Level Architecture ``` ┌─────────────────────────────────────────────────────────────┐ │ CLIENT LAYER │ │ ┌─────────────────┐ ┌──────────────────────────┐ │ │ │ Web (Next.js) │ │ Mobile (Responsive) │ │ │ └────────┬────────┘ └────────────┬─────────────┘ │ └───────────┼──────────────────────────────┼─────────────────┘ │ │ └──────────────┬───────────────┘ │ HTTPS / REST ┌──────────────────────────▼─────────────────────────────────┐ │ API LAYER (FastAPI) │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────────┐ │ │ │ /valuation │ │ /explain │ │ /health │ │ │ └──────┬───────┘ └──────┬───────┘ └──────────────────┘ │ └─────────┼─────────────────┼────────────────────────────────┘ │ │ ┌─────────▼─────────────────▼────────────────────────────────┐ │ SERVICE LAYER (Python) │ │ ┌──────────────────────────────────────────────────────┐ │ │ │ Valuation Engine (Core) │ │ │ │ ┌────────────┐ ┌────────────┐ ┌───────────────┐ │ │ │ │ │ Cleaner │→ │ Feature │→ │ Price │ │ │ │ │ │ Normalizer │ │ Weighting │ │ Estimator │ │ │ │ │ └────────────┘ └────────────┘ └───────┬───────┘ │ │ │ │ │ │ │ │ │ ┌────────────┐ ┌────────────┐ ┌───────▼───────┐ │ │ │ │ │ Explain- │← │ Confidence │← │ Range │ │ │ │ │ │ ability │ │ Scorer │ │ Builder │ │ │ │ │ └────────────┘ └────────────┘ └───────────────┘ │ │ │ └──────────────────────────────────────────────────────┘ │ └──────────────────────────┬─────────────────────────────────┘ │ ┌──────────────────────────▼─────────────────────────────────┐ │ DATA ACCESS LAYER │ │ ┌──────────────────────────────────────────────────────┐ │ │ │ Repository Pattern (abstracted) │ │ │ │ ┌──────────────┐ ┌──────────────┐ ┌───────────┐ │ │ │ │ │ PostgreSQL │ │ Cache (Redis)│ │ S3 / GCS │ │ │ │ │ │ (normalized) │ │ (hot prices) │ │ (raw data)│ │ │ │ │ └──────────────┘ └──────────────┘ └───────────┘ │ │ │ └──────────────────────────────────────────────────────┘ │ └──────────────────────────┬─────────────────────────────────┘ │ ┌──────────────────────────▼─────────────────────────────────┐ │ EXTERNAL DATA SOURCES (Abstracted) │ │ ┌──────────────┐ ┌──────────────┐ ┌──────────────────┐ │ │ │ Classified │ │ Dealer │ │ Market │ │ │ │ Ads (noisy) │ │ Prices │ │ Index (TUIK) │ │ │ └──────────────┘ └──────────────┘ └──────────────────┘ │ └─────────────────────────────────────────────────────────────┘ ``` ### Data Flow ``` 1. INGESTION Raw listings from multiple sources → Normalize schema (brand, model, year, km, price, city, ...) → Store raw in S3/GCS (immutable, timestamped) 2. CLEANING Read raw → apply statistical filters → clean dataset → Store normalized records in PostgreSQL 3. MODELING Clean data → compute feature weights → build price distribution → Cache aggregated statistics in Redis 4. SERVING User input → lookup relevant cohort → compute valuation → Generate explanation → return response ``` ### Key Components | Component | Responsibility | Technology | |-----------|---------------|------------| | **Ingestion Service** | Pull from abstracted sources, normalize schema | Python, Pandas | | **Cleaning Pipeline** | Outlier removal, deduplication, validation | Python, NumPy | | **Valuation Engine** | Core pricing logic (Step 2) | Python, Scikit-learn | | **Explanation Generator** | Human-readable reasoning | Python, templates | | **API Layer** | REST endpoints, validation, rate limiting | FastAPI, Pydantic | | **Frontend** | User interface, responsive | Next.js, Tailwind CSS | | **Cache** | Hot price lookups, session data | Redis | | **Database** | Normalized data, user history | PostgreSQL | ### Data Model (High-Level) ``` listings ├── id (UUID) ├── source (string) ├── brand (string) ├── model (string) ├── year (int) ├── fuel_type (enum) ├── transmission (enum) ├── mileage_km (int) ├── city (string) ├── price_try (numeric) ├── damage_status (enum) ├── ownership_count (int) ├── engine_size_cc (int, nullable) ├── trim_level (string, nullable) ├── color (string, nullable) ├── usage_type (enum, nullable) ├── accident_severity (enum, nullable) ├── listed_at (timestamp) ├── is_outlier (boolean) └── confidence_score (float) valuation_requests ├── id (UUID) ├── input_features (JSONB) ├── result_range (JSONB) ├── confidence (float) ├── created_at (timestamp) └── user_feedback (enum, nullable) ``` ### Key Design Decisions & Justifications - **Hybrid approach (rule-based + statistical)**: Avoids black-box complexity, ensures explainability from day one. ML can be added later for fine-tuning. - **Repository pattern for data access**: Allows swapping data sources without touching core logic. Critical given noisy, volatile market. - **PostgreSQL for normalized data**: Reliable, supports JSONB for flexible features, good for analytical queries. - **Redis for caching**: Price distributions per cohort (brand/model/year/city) are expensive to compute; cache them with TTL. - **FastAPI over Django**: Lightweight, async-ready, auto-generated OpenAPI docs, Pydantic validation aligns with our input schema needs. - **Next.js for frontend**: Server-side rendering for SEO, mobile-first responsive design, easy API integration. ### Assumptions - Data sources are abstracted; we assume a normalized ingestion interface. - Initial data volume: ~500K-1M records (manageable with PostgreSQL). - Currency: Turkish Lira (TRY). All prices stored in TRY. - Inflation adjustment: We will use TUIK (Turkish Statistical Institute) automotive price index as a macro-level correction factor. - City effects: 81 provinces, but we will group into 7 regions + Istanbul/Ankara/Izmir as special cases. ### Risks & Mitigations | Risk | Mitigation | |------|-----------| | **Market manipulation** | Statistical outlier removal, median-based estimators, cross-source validation | | **Inflation volatility** | Time-decay weighting on listings, index-based adjustment | | **Sparse data for rare models** | Fallback to brand-level aggregates, confidence score reflects data sparsity | | **Feature mismatch** | Normalize inputs, use nullable fields, provide sensible defaults | --- **Do you want to proceed to Step 2 – Valuation Logic Design?**

Türkische Fahrzeugbewertungsplattform Builder

Prompt-InhaltSpeichern

Verwendung

Referenzen

Diskussion