DIE-Ops is an end-to-end Decision Intelligence prototype designed to help enterprises convert customer analytics into actionable, budget-constrained decisions.
Rather than stopping at predictions (churn scores, CLTV, forecasts), DIE-Ops focuses on the harder and more valuable problem:
Who should we act on, why, under a fixed budget, and what business impact should we expect?
This project reflects how real-world enterprises operationalize AI to drive measurable ROI, bridging the gap between data science and business outcomes.
Most organizations already have churn models, customer segmentation, and dashboards. Yet they still struggle to answer basic operational questions:
- Which customers should we target right now?
- How much should we spend?
- What return can we realistically expect?
- Why were these customers chosen?
The gap is not modeling—it is decision-making. DIE-Ops exists to bridge analytics and action.
Given a fixed marketing budget and cost per action, which customers should we target to maximize incremental business impact—and why?
DIE-Ops combines three complementary decision lenses to move beyond simple propensity scoring:
- Churn (Risk): Estimates the probability that a customer will disengage.
- Answers: "Who is at risk?"
- CLTV (Value): Estimates the future economic value of a customer using repeat-purchase behavior (BG/NBD + Gamma-Gamma).
- Answers: "Who is worth saving?"
- Uplift (Impact): Estimates the incremental change caused by taking an action (e.g., an offer).
- Answers: "Who will actually change behavior if we act?"
DIE-Ops combines all three into ROI-aware decisions.
[Customer Data: CSV / SQL] | v [Ingestion & Validation] | v [Feature Engineering] | v [ML Layer]
- Churn Model (LightGBM)
- CLTV Model (Lifetimes)
- Uplift Model (Two-Model Approach) | v [Decision Engine]
- ROI-based Optimizer (Greedy Selection)
- Budget & Cost Constraints | v [Explainability Layer]
- SHAP-based narratives
- Scenario simulation | v [Delivery]
- FastAPI endpoints (Scoring & Decisions)
- Flask + HTMX UI (Interaction)
- Executive-ready HTML reports & CSV exports
For a given campaign configuration, the system outputs:
- Prioritized Target List: A specific list of customers sorted by ROI.
- Financial Projections: Expected incremental revenue and total campaign ROI.
- Reason Codes: SHAP-derived explanations for why a customer was selected.
- Strategy Reports: Automated, executive-ready HTML reports for stakeholders.
- Operational Exports: CSV files ready for CRM ingestion.
- Enterprise Data Layer: Deterministic ingestion with schema validation and checksums for reproducibility.
- Multi-Model Intelligence: Integrated Risk, Value, and Impact modeling.
- Decision & Optimization: ROI-aware selection that explicitly excludes negative-ROI actions ("Sleeping Dogs").
- Explainability: SHAP-based explanations converted into business-friendly narratives.
- Full-Stack Mindset: FastAPI for serving, Flask/HTMX for UI, and CI-ready Docker structures.
DIE-Ops demonstrates:
- Decision-first thinking, not just model-first thinking.
- Understanding of how AI is operationalized in enterprise environments.
- The ability to design systems that connect Data → Decisions → Outcomes.
- Alignment with high-level analytics consulting and AI transformation work.
# Generate synthetic dataset
python src/ingest/ingest.py# Train the three-pillar intelligence layer
python -m src.models.train_churn
python -m src.models.train_uplift
python -m src.models.train_cltv# Start the FastAPI Backend
uvicorn src.api.main:app --reload
# In a new terminal, start the Flask UI
python src/frontend/server.py
Access the dashboard at http://localhost:5001.- Online A/B validation of predicted uplift.
- Campaign monitoring and automated drift detection.
- Integration with production CRM platforms (e.g., Salesforce, HubSpot).
- Multi-action policy optimization (Choosing the best offer, not just any offer).
This project was built to demonstrate how applied AI systems should be designed when the goal is better decisions, not just better models.