Vinctum Aegis
From churn risk to an economically grounded retention decision.
Combines risk, explanations, intervention cost and expected value in one workflow.
- For
- Subscription and customer lifecycle teams
- Engine
- Python · XGBoost · SHAP
- Source code
- Closed · shown during a pilot
- Buying
- Written, fixed-price pilot
What is Vinctum Aegis?
Vinctum Aegis is a decision system for customer churn. For each customer it computes a calibrated churn risk, the reasons behind it and the cost of possible interventions, then says which intervention has a positive expected value. It is a working, closed-source MVP and claims no verified customer outcome.
The same 78% risk, two opposite decisions.
Same risk score
Same risk score
Representative assumptions: the email lowers risk by 4%, outreach by 25%, the discount by 12%. The real effect is only known from a controlled measurement.
How it works
From the product's own documentation. Each number says where and on which version it was measured.
One feature pipeline
Polars preprocessing persists feature state, so training and serving see the same features.
Calibrated risk
Gradient boosting models emit calibrated probabilities: '78%' actually means close to 78%.
Why, and what would change it
SHAP explains the risk; DiCE shows which change would flip the outcome.
An economic decision
Risk, lifetime value, margin and intervention cost are weighed together. Nothing is recommended below the threshold.
- T-learner
- The uplift engine requires a real treatment/control column, and labels results as simulation otherwise.
- Operations dashboard
- Next.js: SHAP, DiCE, uplift, policy, drift and customer drill-down in one place.
- JWT + RBAC
- Tenant-scoped access, rate limits and an audit log that redacts sensitive fields.
- Closed source
- The code is shown during a pilot, alongside your own data.
Stack
- Python
- Polars
- XGBoost / LightGBM / CatBoost
- SHAP
- DiCE
- Litestar
- Next.js
What it does not claim
The IBM Telco dataset used in development cannot prove causal uplift. Whether an intervention works is measured only with real campaign data and a control group.
Current boundary
The current demo uses representative data. Real impact requires customer data and a control group.
INTERACTIVE EXAMPLE / REPRESENTATIVE DATA
The same risk can call for a different decision.
Change churn risk and monthly customer value. Expected value is recalculated from each intervention’s assumed effect and cost.
This demonstrates decision logic. Effects and costs are representative assumptions; actual impact requires suitable customer data and controlled measurement.
What a pilot covers
A sample starting scope. Final deliverables and acceptance criteria are fixed in the written proposal.
- Question
- Which intervention is worth making for each customer?
- Likely input
- Customer history, churn labels and intervention costs
- Deliverables
- Inspectable risk and reason view
- Priority list grounded in cost and expected value
- Measurement plan with a control group
Before you decide
Can I buy this product today?
A paid pilot and setup proposal can be prepared. Scope, data access, deliverables, timeline and a fixed price are agreed in writing after a discovery call.
Why is there no price?
Data volume, integration and success criteria differ from job to job. After the call you get a fixed-price written proposal; no paid work starts without your approval.
Do I need to upload live data here?
No. This site has no file upload or accounts. Data access, retention and permissions are set only in the pilot agreement.
What is the first step?
Choose Vinctum Aegis in the form and describe the problem you need to solve. The data and pilot boundaries are then set together.
More in this discipline
- LuceThemisRust · Python
Audits signal quality in quantitative research.
- Calybris CoreRust · Python
Decision infrastructure with replayable proof.
- Churn Report KitPython · XGBoost
Churn analysis with an actionable retention queue.