Plataforma de Valoración de Coches Turcos

## Step 1 – Product & System Design ### High-Level Architecture We propose a **modular, layered architecture** with clear separation between data, intelligence, and presentation: ``` ┌─────────────────────────────────────────────────────────┐ │ PRESENTATION LAYER │ │ Web (React/Next.js) + Mobile (Responsive PWA) │ └──────────────────────────┬──────────────────────────────┘ │ REST/JSON ┌──────────────────────────▼──────────────────────────────┐ │ API LAYER (FastAPI) │ │ - Input validation & normalization │ │ - Orchestrates valuation pipeline │ │ - Serves explainability payloads │ └──────────────────────────┬──────────────────────────────┘ │ ┌──────────────────────────▼──────────────────────────────┐ │ VALUATION CORE (Python) │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ Data │ │ Statistical │ │ Explain- │ │ │ │ Ingestion │→ │ Filtering │→ │ ability │ │ │ │ & Cleaning │ │ & Modeling │ │ Generator │ │ │ └─────────────┘ └─────────────┘ └─────────────┘ │ └──────────────────────────┬──────────────────────────────┘ │ ┌──────────────────────────▼──────────────────────────────┐ │ DATA SOURCES (Abstracted) │ │ - Classified listings (noisy) │ │ - Dealer price feeds │ │ - Historical transaction records │ │ - Macroeconomic indicators (inflation, FX) │ └──────────────────────────────────────────────────────────┘ ``` --- ### Key Components | Component | Responsibility | Notes | |-----------|---------------|-------| | **Data Ingestion Abstraction** | Unified interface for multiple data sources | Each source adapter normalizes to a common schema; no direct scraping unless explicitly allowed | | **Cleaning & Normalization Engine** | Removes outliers, detects manipulation, standardizes units | Uses robust statistics (IQR, MAD) rather than mean/std | | **Feature Weighting Module** | Computes contribution of each input variable | Rule-based + statistical hybrid; transparent weights | | **Price Distribution Modeler** | Builds price distribution per car segment | Uses percentiles (P10, P50, P90) for range output | | **Explainability Layer** | Generates human-readable reasoning | Maps feature contributions to natural language explanations | | **API Gateway** | Validates inputs, orchestrates pipeline, formats responses | FastAPI with Pydantic schemas | | **Frontend App** | User input form, results dashboard, explanation view | React/Next.js, mobile-first | --- ### Data Flow ``` User Input (car features) │ ▼ [1] Input Validation & Normalization │ ▼ [2] Data Source Fetch (abstracted, cached) │ ▼ [3] Cleaning & Outlier Removal │ ▼ [4] Feature Weighting & Price Modeling │ ▼ [5] Range Generation (min/fair/max) + Confidence Score │ ▼ [6] Explainability Generation │ ▼ [7] API Response → Frontend Display ``` --- ### Key Design Decisions & Justifications **Decision 1: Statistical filtering over raw averages** - *Why:* Turkish listing prices are highly volatile and manipulated. Robust estimators (median, trimmed mean, percentile-based) resist outliers better than mean-based approaches. **Decision 2: Hybrid rule-based + statistical model (not pure ML)** - *Why:* Explainability is a core product requirement. Black-box models would undermine user trust. We start with transparent rules and add ML only if it improves accuracy without sacrificing explainability. **Decision 3: Modular pipeline with clear interfaces** - *Why:* Allows swapping data sources, adjusting weighting logic, or upgrading models independently without breaking the system. **Decision 4: Mobile-first responsive design** - *Why:* Target users in Turkey predominantly access via mobile. PWA approach avoids separate native app development initially. **Decision 5: Caching layer for data sources** - *Why:* Reduces latency and dependency on external sources. Data freshness is balanced with performance. --- ### Assumptions 1. **Data sources** are abstracted and may be synthetic initially; real integration requires explicit permission. 2. **Macroeconomic factors** (inflation, FX rates) are treated as external inputs that can be updated periodically, not real-time. 3. **City-based adjustments** are derived from regional price indices, not raw listing averages. 4. **Confidence score** is based on data volume, variance, and feature completeness. 5. **No user accounts** required for basic valuation; optional for saving history. --- ### Risks & Mitigations | Risk | Mitigation | |------|-----------| | Data quality issues | Robust statistical filtering; multiple source cross-validation | | Market volatility | Regular model recalibration; confidence score reflects uncertainty | | User distrust of estimates | Transparent explainability; show data basis and limitations | | Scope creep | Strict MVP focus; iterative feature addition | --- Do you want to proceed to **Step 2 – Valuation Logic Design**?

Plataforma de Valoración de Coches Turcos

Contenido del PromptGuardar

Uso

Referencias

Discusión