SaaSGuard

Production-grade B2B SaaS churn and compliance risk prediction platform.

---

What is SaaSGuard?

SaaSGuard ingests product usage, CRM, and support data to answer one high-value question:

Which customers will churn in the next 90 days — and why?

It combines survival analysis, XGBoost classification, SHAP explainability, and an AI executive summary layer to give Customer Success teams the right signal at the right time.

Business impact: Reducing churn by 1% on $200M ARR = $2M+ revenue saved annually.

---

The Problem (Voice-of-Customer)

Pain Point	Source	Impact
"Initial onboarding could be more guided"	G2 Reviews	20–25% churn in first 90 days
"More manual input than expected"	Forrester	Integration abandonment → churn
"Very difficult to reach them when you have a problem"	Industry surveys	Reactive CS → preventable churns
"80% of B2B buyers switch when expectations aren't met"	SaaS industry reports	Poor activation → early exit

SaaSGuard attacks each of these by surfacing the right signal early enough to act.

---

Quick Demo

git clone https://github.com/joewynn/SaaSGuard
cd SaaSGuard
cp .env.example .env
docker compose --profile dev up -d

Service	URL	Purpose
FastAPI	localhost:8000/docs	Prediction & customer API
Superset	localhost:8088	BI dashboard (Customer 360)
JupyterLab	localhost:8888	EDA notebooks
MkDocs	localhost:8001	This documentation site

---

Architecture at a Glance

graph LR
    subgraph "Data Sources"
        A[usage_events]
        B[customers]
        C[support_tickets]
        D[risk_signals]
    end

    subgraph "Transformation"
        E[(DuckDB)]
        F[dbt Models]
    end

    subgraph "Domain Layer (DDD)"
        G[customer_domain]
        H[usage_domain]
        I[prediction_domain]
        J[gtm_domain]
    end

    subgraph "Delivery"
        K[FastAPI :8000]
        L[Superset :8088]
        M[MkDocs :8001]
    end

    A & B & C & D --> E
    E --> F --> G & H & I & J
    I --> K --> L
    K --> M

Full DDD diagram with request flow → Architecture.

---

Engineering principles

Four constraints shaped every architectural decision, documented formally in the ADRs above.

Auditability over convenience. Feature engineering lives in dbt SQL, not Python. Every risk tier surfaced in a dashboard is reproducible from mart_customer_risk_scores alone — no Python runtime required. CS and Compliance teams can audit a score without reading model code.

Calibration is non-negotiable. CalibratedClassifierCV (isotonic, cv=5) ensures a predicted probability of 0.72 corresponds to a 72% historical churn rate in that decile.

Grounded AI, not generative AI. The executive summary layer is constrained to DuckDB-verified facts. The GuardrailsService checks every output against a feature whitelist and the model's probability output. Confidence degrades 0.2 per violation.

Zero-dependency domain layer. src/domain/ has no file I/O, no database calls, and no HTTP dependencies. Every prediction path is fully unit-testable with injected fakes. The 153-test suite completes in under 8 seconds locally.