An end-to-end Advanced Analytics consulting engagement built to BCG X Delivery team standards. From raw business problem to executive recommendations — production Python, not notebooks.

NovaMart is a Fortune 500 specialty retail chain operating 800+ stores across North America with $12B in annual revenue. Over 18 months, net profit margin declined from 8.2% to 5.7% — a $300M annualized impact. Leadership could not determine whether the driver was pricing, product mix, store operations, customer behavior, or marketing inefficiency.

This project delivers the analytics engagement that answers that question.

┌─────────────────────────────────────────────────────────────────┐
│                  Streamlit Dashboard  (port 8501)               │
│   Executive KPIs · EDA · Simulations · Recommendations ·        │
│   Statistical Analysis · Model Performance · AI Copilot        │
├─────────────────────────────────────────────────────────────────┤
│                   FastAPI Backend  (port 8000)                  │
│   /api/v1/data · /simulation · /recommendations · /copilot     │
├──────────────────┬──────────────────┬───────────────────────────┤
│  Analytics Layer │   ML Layer       │  GenAI Copilot            │
│  EDA · Hypothesis│  Churn · Store   │  Claude API (tool-use)    │
│  RFM · Elasticity│  MMM · Elasticity│  7 analytics tools        │
│  A/B Testing     │  SHAP explained  │  Grounded, no hallucination│
├──────────────────┴──────────────────┴───────────────────────────┤
│         Feature Store  (Parquet, data/features/)                │
├─────────────────────────────────────────────────────────────────┤
│         Data Layer — Synthetic NovaMart Data                    │
│   7 tables · 1.38M rows · 36 months · Reproducible (seed=42)  │
└─────────────────────────────────────────────────────────────────┘

These are not illustrative — they are computed from the generated dataset:

• Margin decline confirmed: Gross margin fell from 42.1% → 31.8% over 36 months (−24.5% relative). Statistically significant OLS trend (p<0.001, R²=0.91). Root cause: COGS inflation beginning Month 13, growing ~0.75% per month with no corresponding price adjustment.

• Silver tier churn crisis: Silver loyalty customers show materially lower purchase activity from Month 25 onward, following a loyalty fee increase. Confirmed via Welch's t-test (p<0.001, Cohen's d=0.72 — large effect). Revenue at risk: $8–12M annually.

• Store concentration risk: Cluster A stores (17% of portfolio, 140 stores) generate 43% of total profit. Cluster C stores (25% of portfolio, 202 stores) generate only 9%. Top 10% of stores account for 47% of revenue — significant concentration risk.

• Private label margin advantage: Private label products earn a 2.30x price-to-cost multiplier vs 1.60x for national brands — 44% more margin per unit. Private label penetration is declining despite this advantage. Five-point share increase = ~$4.2M incremental annual profit.

• Marketing ROI gap: Digital channels deliver 3.0x revenue per dollar in urban stores vs TV at 0.8x. Pattern reverses in rural stores (TV = 2.5x, digital = 0.8x). Current allocation ignores this segmentation.

# 1. Clone
git clone https://github.com/your-username/BCG-X-Analytics-Accelerator.git
cd BCG-X-Analytics-Accelerator

# 2. Install
pip install -e ".[dev]"

# 3. Configure environment
cp .env.example .env
# Optional: add ANTHROPIC_API_KEY to enable the AI Copilot

# 4. Generate synthetic NovaMart data (~20 seconds)
make generate-data

# 5. Run data audit
python scripts/run_audit.py

# 6. Run exploratory analysis
python scripts/run_eda.py

# 7. Train ML models (optional, ~5 minutes)
make train-models

# 8. Generate recommendations
python scripts/generate_recommendations.py

# 9. Generate executive summary report
python scripts/generate_slides.py
# → data/outputs/NovaMart_Executive_Summary.md

# 10. Launch Streamlit dashboard
make run-dashboard
# → http://localhost:8501

# 11. Launch FastAPI backend (separate terminal)
make run-api
# → API docs: http://localhost:8000/api/docs

# Full Docker stack (all services at once)
make docker-up

The copilot implements a full agentic loop using Anthropic Claude's native tool-use API — not a chatbot with hardcoded answers. Claude receives the user's question alongside 7 analytics tool definitions, decides which to call, executes them against live NovaMart parquet data, and synthesises a consulting-quality answer grounded in real figures. The loop repeats up to 5 iterations so Claude can chain tools for complex, multi-part questions.

Why tool-use instead of RAG? The analytics layer returns structured JSON (exact KPIs, p-values, model metrics) — vector similarity retrieval would introduce noise on data that is already perfectly queryable. Tool-use gives Claude precise numbers, not approximate matches.

Every response shows which tools were called, giving full auditability. Claude cannot cite a figure it didn't retrieve — if data isn't available it says so rather than hallucinating.

The 7 tools: get_financial_summary · get_churn_analysis · get_store_performance · get_marketing_roi · get_pricing_analysis · get_recommendations · simulate_scenario

To enable: set OPENROUTER_KEY or ANTHROPIC_API_KEY in .env.

Four business what-if scenarios, each computing real financial impact from the data:

This project demonstrates end-to-end analytics consulting delivery: from business problem framing through statistical rigor, production ML, and executive communication.

Every analytical output is framed around a business question and concludes with a recommended action — not a metric. The architecture mirrors BCG X Delivery standards: modular Python packages, tracked experiments, SHAP explainability, and a decoupled FastAPI/Streamlit stack.

The GenAI Copilot uses the tool-use pattern so responses are always grounded in the analytics layer — demonstrating responsible AI design, not just LLM capability.

MIT