A production-grade Business Intelligence microservice that connects to an ERP database and exposes ML-powered analytics — product scoring, employee performance evaluation, sales forecasting, recommendation engine, and market basket analysis — through a clean REST API.
Built with Flask + scikit-learn + Facebook Prophet and deployed via Docker + Nginx.
| Module | Technique | Endpoint |
|---|---|---|
| 📦 Product Performance Analysis | K-Means clustering + weighted scoring | POST /api/product-analysis |
| 👥 Employee Performance Evaluation | K-Means + Isolation Forest anomaly detection | POST /api/analyze |
| 📈 Sales Forecast | Linear Regression with train/test split | GET /api/sales-forecast |
| 🔮 Prophet Forecast | Facebook Prophet (seasonality + trend) | GET /api/prophet-forecast |
| 🛒 Product Recommendations | KNN content-based + collaborative filtering + trending | POST /api/recommendations/products |
| 🧺 Market Basket Analysis | Apriori algorithm (frequent itemsets + association rules) | GET /api/product-combos |
Sample dashboards rendered from each endpoint (synthetic data; generators in docs/tools/*.py, full write-ups in docs/models/):
Module 1 — Product Performance Analysis
Module 2 — Employee Performance Evaluation
Module 3 — Sales Forecast (Linear Regression)
Module 4 — Sales Forecast (Facebook Prophet)
Module 5 — Product Recommendations
Module 6 — Market Basket Analysis (Apriori)
┌─────────────┐ ┌──────────────┐ ┌───────────────┐
│ Client │────▶│ Nginx │────▶│ Flask App │
│ (Dashboard) │ │ (Reverse │ │ (BI Service) │
└─────────────┘ │ Proxy) │ └───────┬───────┘
└──────────────┘ │
┌──────▼───────┐
│ MariaDB │
│ (ERP DB) │
└──────────────┘
- 🌐 Nginx — reverse proxy, request routing, timeout tuning
- 🧠 Flask App — 6 route blueprints, lazy model loading, joblib persistence
- 🗄️ MariaDB — shared external ERP database (not managed by this container)
- 💾 Trained models saved as
.joblibfiles underml_models/
- Backend: 🐍 Python 3.11, ⚗️ Flask 3.0, Flask-SQLAlchemy, Flask-CORS
- ML/DL: 📐 scikit-learn, pandas, numpy, mlxtend (Apriori), Facebook Prophet
- Infrastructure: 🐳 Docker, Docker Compose, Nginx
- Storage: 🗄️ PyMySQL, joblib (model serialization)
- Docker & Docker Compose
- Access to a MariaDB/MySQL ERP database
git clone <repo-url>
cd bi-systemEdit flask-app/.env to point to your database:
DB_HOST=your_mariadb_host
DB_PORT=3306
DB_NAME=laravel_db
DB_USER=root
DB_PASSWORD=rootEnsure the network erp-system_default exists (or change docker-compose.yml):
docker network create erp-system_defaultdocker compose up -d| Service | Port | URL |
|---|---|---|
| Flask API | 5001 |
http://localhost:5001 |
| Nginx | 5002 |
http://localhost:5002 |
curl http://localhost:5001/
# {"status":"ok","service":"BI System","endpoints":[...]}Analyzes all active products using K-Means clustering + weighted scoring (revenue, stock ratio, sales velocity, profit margin).
// Response
{
"status": "success",
"data": [{"product_id": 1, "name": "...", "performance_score": 85.3, "performance_tier": "High", ...}],
"summary": {"total_products": 50, "avg_score": 62.4, "high_tier": 12, ...},
"model_version": "2.0"
}Evaluates employee performance based on task completion rates, task velocity, and identifies anomalies (Isolation Forest).
Linear regression forecast for next month's revenue. Returns historical monthly aggregated sales + prediction with confidence interval.
Facebook Prophet forecast. Parameters: ?periods=12 (1–60 months). Returns historical fit, future forecast, trend decomposition, and summary metrics.
Multiple recommendation strategies via JSON body:
{
"strategy": "popular|trending|content_based|collaborative|personalized",
"product_id": 123, // for content_based
"user_id": 456, // for collaborative / personalized
"limit": 10
}Manually retrain the recommendation models (KNN + collaborative filtering).
Runs Apriori algorithm on completed sales transactions. Returns frequent itemsets and association rules (lift, confidence, support).
bi-system/
├── docker-compose.yml # Service orchestration
├── nginx/
│ └── default.conf # Reverse proxy config
├── flask-app/
│ ├── Dockerfile # Python 3.11 + Prophet + dependencies
│ ├── requirements.txt
│ ├── .env # DB credentials (not committed)
│ ├── app.py # Flask app factory, blueprint registration
│ ├── models/
│ │ └── __init__.py # All SQLAlchemy ORM models (15 tables)
│ ├── routes/
│ │ ├── __init__.py
│ │ ├── employee_performance.py # K-Means + Isolation Forest
│ │ ├── product_analysis.py # K-Means + weighted scoring
│ │ ├── product_recommendations.py # KNN + collaborative + trending
│ │ ├── product_combos.py # Apriori association rules
│ │ ├── sales_forecast.py # Linear Regression
│ │ └── prophet_forecast.py # Facebook Prophet
│ └── ml_models/ # Trained .joblib models (gitignored)
├── docs/ # Module documentation, dashboard images, HTML export
│ ├── models/ # In-depth per-module write-ups (Markdown)
│ ├── images/ # Generated demo dashboard PNGs
│ ├── html/ # HTML export of the Markdown docs
│ └── tools/ # Demo dashboard generators + md_to_html.py
├── README.md
├── AGENTS.md # AI agent onboarding guide
└── case-study.md # Comprehensive case study
All endpoints return "model_version": "2.0". Models retrain on every request (except recommendation models which use lazy loading). Trained models are persisted to ml_models/*.joblib.
MIT