A production-style internal analytics platform giving Hawkins management a single pane of glass over dealer performance, regional demand, inventory health, service quality, and competitive positioning — replacing scattered Excel reports with an automated, ML-augmented system.
Built by Alok Deep · Portfolio project for the Hawkins Cookers T/Systems role.
Most candidates for an in-house IT/Systems role hand over a generic CRUD app or a Kaggle notebook. Neither answers the question a hiring manager actually has: can you build something our business would use?
This project tries to. It models Hawkins' actual operating reality — 9,379 dealers, three plants, sixteen brand lines, the FY25 growth recovery — and builds the kind of internal tool an in-house team would ship to replace email-attached Excel reports. Synthetic data, real architecture: a 1.05M-row SQLite warehouse, a FastAPI backend with 30+ endpoints, a React + TypeScript frontend, and an ML layer doing demand forecasting, anomaly detection, and dealer segmentation.
Why ML and SQL, not one or the other? Because the questions IT inherits from the business mix both — "what will Bigboy demand look like in Q3?" needs a model; "which dealers slipped 30% MoM?" needs a query. The platform handles both, and the JD calls for both.
 |
 |
| Home — Entry point with module navigation | Executive Overview — Revenue, KPIs, YoY trends |
 |
 |
| GIS Distribution — Pan-India choropleth, white-space scoring | Dealer Performance — Scoring, movers, cohort drill-down |
 |
 |
| Demand Forecasting — Per-SKU SARIMA with confidence bands | Anomaly Detection — Z-score + Isolation Forest dual-method |
 |
 |
| Service Analytics — Warranty claims, quality-risk signals | Competitive Intelligence — Pricing gap vs Prestige, TTK, Butterfly |
| JD requirement | Where it lives in the project |
|---|---|
| SQL · Python | data/hawkins.db (1.05M rows, 10 indexes, 5 analytical views) + ETL pipeline |
| Designing reports | 7 dashboard pages, drill-down filters via shared FiltersContext |
| Implementing automation | run_pipeline.py — one-command generate → ETL → train; start.bat to launch the stack |
| ML / AI | SARIMA per-SKU forecasts · Isolation Forest anomalies · RFM K-Means segmentation |
| Consumer / customer analytics | Dealer scoring, dormancy alerts, cohort analysis |
| GIS mapping | GISDistribution.tsx — choropleth + state drill-down + white-space scoring |
| In-house system development | Modular routers, shared context, documented architecture |
| Modern BI stack | FastAPI · React + TypeScript · Tailwind · Recharts · react-leaflet · TanStack Query |
All data is synthetic, but every dimension is anchored to publicly verified Hawkins reality. This is the difference between a "demo dataset" and a model the business would recognise.
| Anchor | Real Hawkins Fact | This Project |
|---|---|---|
| Revenue | ₹1,030 Cr (FY24), ₹1,194 Cr (FY25) | ₹917 Cr over 3 years simulated |
| Dealers | 9,379 authorised dealers (ICRA 2024) | 1,900 (5× sampled for performance) |
| Service centres | ~700 across India / Nepal / Bhutan | 140 |
| Plants | Thane (MH), Hoshiarpur (PB), Sathariya (UP) | All 3 modelled with capacities |
| Brand lines | Classic, Contura, Futura, Stainless, Hevibase, Bigboy, Miss Mary, etc. | All 16+ lines with realistic SKU mix |
| Market share | ~32% Indian pressure cooker segment | Reflected in distribution patterns |
| Growth | FY23: +5%, FY24: +2%, FY25: recovery | Encoded in 3-year YoY trend |
| Seasonality | Diwali, Akshaya Tritiya, wedding-season peaks | Coded into transaction date weights |
Sources: Hawkins annual reports, ICRA rating rationale (2024), company website.
┌──────────────────────────────────────────────────────────────────┐
│ DATA GENERATION LAYER │
│ scripts/data_generation/ → data/raw/*.csv │
│ Products · Geography · Dealer network · Sales · Aux tables │
└───────────────────────────┬──────────────────────────────────────┘
▼
┌──────────────────────────────────────────────────────────────────┐
│ ETL / DATABASE LAYER │
│ scripts/etl/load_to_sqlite.py → data/hawkins.db (SQLite) │
│ 10 indexes · 5 analytical views · 1 materialised summary │
└───────────────────────────┬──────────────────────────────────────┘
▼
┌─────────────────┴──────────────────┐
▼ ▼
┌──────────────────┐ ┌─────────────────────────────┐
│ ML LAYER │ │ API LAYER │
│ │ │ backend/ (FastAPI) │
│ SARIMA forecasts│ │ 7 routers · 30+ endpoints │
│ Isolation Forest│ │ SQLite read-only · CORS │
│ RFM + K-Means │ └──────────────┬──────────────┘
│ → models/*.pkl │ ▼
└──────────────────┘ ┌─────────────────────────────┐
│ UI LAYER │
│ frontend/ (React + Vite) │
│ TypeScript · Tailwind CSS │
│ Recharts · react-leaflet │
│ TanStack Query │
└─────────────────────────────┘
Design choices worth calling out:
- SQLite, not Postgres. A reviewer can clone this repo and have a working warehouse in two minutes. The SQL is portable; the deployment story is zero-friction. For a production rollout the swap to Postgres or MSSQL is a connection-string change.
- Read-only DB connection in the API. The dashboard cannot mutate the warehouse — separation of concerns enforced at the connection layer, not just by convention.
- Shared
FiltersContexton the frontend. State (date range, region, brand line) lives in one place; every page subscribes. Avoids the prop-drilling soup that kills these dashboards. - ML artifacts as files, not services.
forecast_*.pkl,iso_anomalies.parquet— the API loads them at startup. Keeps the runtime simple and the retraining pipeline decoupled.
| Model | Library | What it answers | Output |
|---|---|---|---|
| SARIMA (per-SKU) | statsmodels |
"What will demand for Futura 3L look like next quarter?" | Point forecast + 95% confidence band |
| Isolation Forest | scikit-learn |
"Which transactions look structurally weird?" — multivariate, picks up patterns z-score misses | Anomaly score per row |
| Z-score (baseline) | numpy |
"Which transactions are statistical outliers on a single dimension?" | Pairs with Isolation Forest in dual-method UI |
| RFM + K-Means | scikit-learn |
"How do we segment 1,900 dealers — champions, dormant, at-risk, new?" | Cluster label + RFM tuple per dealer |
Methodology, train/test splits, and metrics are documented in docs/ML_METHODOLOGY.md.
Prerequisites: Python 3.10+ · Node.js 18+
# 1. Create and activate virtual environment
python -m venv venv
venv\Scripts\activate # Windows
# source venv/bin/activate # macOS / Linux
# 2. Install Python dependencies
pip install -r backend/requirements.txt
pip install -r scripts/requirements.txt
# 3. Generate data, build the database, and train ML models (~4 min)
python run_pipeline.py
# Faster rebuild without retraining (~2 min):
python run_pipeline.py --skip-ml
# 4. Install frontend dependencies (first time only)
cd frontend && npm install && cd ..
# 5. Launch both servers
start.bat| Service | URL |
|---|---|
| Frontend (React) | http://localhost:5173 |
| Backend (FastAPI) | http://localhost:8000 |
| API docs (Swagger) | http://localhost:8000/docs |
hawkins-distribution-intelligence-platform/
├── backend/ # FastAPI REST API
│ ├── routers/
│ │ ├── anomalies.py # Z-score + Isolation Forest endpoints
│ │ ├── competitive.py # Pricing gap vs competitors
│ │ ├── dealers.py # Scoring, movers, cohorts
│ │ ├── executive.py # KPIs, revenue trends
│ │ ├── forecasting.py # SARIMA forecast endpoints
│ │ ├── gis.py # Map data, whitespace analysis
│ │ └── service.py # Warranty claims, quality risk
│ ├── db.py # SQLite read-only connection helper
│ ├── main.py # FastAPI app + CORS
│ └── requirements.txt
│
├── frontend/ # React + TypeScript (Vite)
│ ├── public/
│ │ ├── logo.png
│ │ └── screenshots/ # Dashboard screenshots for README
│ └── src/
│ ├── api/client.ts # Axios base client
│ ├── components/ # Layout, Sidebar, ChartCard, KPICard, etc.
│ ├── context/FiltersContext.tsx # Shared filter state across pages
│ └── pages/ # One file per dashboard module
│ ├── Home.tsx
│ ├── Executive.tsx
│ ├── GISDistribution.tsx
│ ├── DealerPerformance.tsx
│ ├── Forecasting.tsx
│ ├── AnomalyDetection.tsx
│ ├── ServiceAnalytics.tsx
│ └── CompetitiveIntel.tsx
│
├── scripts/ # Data pipeline + ML training
│ ├── data_generation/ # Synthetic data generators (5 scripts)
│ ├── etl/ # CSV → SQLite + view optimisation
│ ├── ml/ # SARIMA · Isolation Forest · RFM K-Means
│ └── requirements.txt # Pipeline-only Python deps
│
├── data/
│ ├── raw/ # Generated CSVs (11 files, 1.05M+ rows)
│ ├── processed/ # ETL output staging
│ ├── hawkins.db # SQLite database (gitignored — run pipeline)
│ └── india_states.geojson # State boundary polygons for choropleth
│
├── models/ # Trained ML artifacts (gitignored)
│ ├── forecast_*.pkl # Per-SKU SARIMA models
│ ├── forecast_index.json
│ ├── iso_anomalies.parquet # Isolation Forest results
│ └── zscore_anomalies.parquet
│
├── docs/
│ ├── ARCHITECTURE.md
│ ├── DATA_DICTIONARY.md
│ ├── ML_METHODOLOGY.md
│ ├── DEMO_SCRIPT.md
│ ├── HANDOFF.md
│ └── INTERVIEW_PREP.md
│
├── .gitignore
├── run_pipeline.py # One-command: generate → ETL → ML training
└── start.bat # Launch backend + frontend together
| Layer | Choice | Why |
|---|---|---|
| Database | SQLite | Zero-setup; portable; reviewer can run the project without provisioning anything |
| Backend | FastAPI + Pydantic | Async-ready, auto-generated Swagger docs, type-safe schemas |
| Frontend | React + TypeScript (Vite) | Type-safe component contracts; Vite for fast HMR during development |
| Styling | Tailwind CSS | Avoids the global-stylesheet rot that kills these dashboards |
| Charts | Recharts | Composable, declarative, plays well with React state |
| Maps | react-leaflet | Open-source choropleth with full control — no Mapbox token dependency |
| State | TanStack Query | Caching + invalidation for API calls; takes pressure off useEffect |
| ML | statsmodels · scikit-learn | Proven, well-documented, easy to hand off |
| Doc | Purpose |
|---|---|
docs/ARCHITECTURE.md |
System design, data flow, deployment notes |
docs/DATA_DICTIONARY.md |
Every field in every table |
docs/ML_METHODOLOGY.md |
Model choices, training, evaluation |
docs/DEMO_SCRIPT.md |
5-minute walkthrough for the live demo |
docs/HANDOFF.md |
What a successor would need to maintain this |
docs/INTERVIEW_PREP.md |
STAR answers, likely questions |
Educational / portfolio project. Not affiliated with Hawkins Cookers Ltd. All facts about Hawkins are drawn from publicly available sources (annual reports, ICRA ratings, company website).
Alok Deep — Full-stack developer (MERN) building toward data science / analytics roles.