Multi-Tenant SaaS Analytics Platform with Forecasting Engine

Business outcome. Product, operations, and revenue teams get live KPIs, trustworthy forecasts, and automatic anomaly callouts in one place — without waiting for the monthly BI export. Because the prediction engine is deterministic TypeScript and the LLM only writes the narrative, every number on the dashboard is reproducible and auditable. GraphQL exposes the same engine to downstream apps and pipelines, so the dashboard isn't a dead end.

Built for product and ops teams that need forecasts and anomaly callouts they can defend in a board meeting. The type system keeps the LLM out of the math, so the AI layer can be swapped, retuned, or removed without touching the engine that earns the numbers.

Pipeline

What's built

• Synthetic SaaS dataset. Seeded faker producing 365 days of metrics with linear growth, weekly seasonality, three marketing-spike events, a churn dip, and Gaussian noise. Output is byte-identical across runs.
• Prediction engine. Pure TypeScript, zero deps. Linear-regression forecasting, z-score anomaly detection over a 14-day trailing window, R²-blended-with-CV confidence scoring.
• GraphQL Yoga API. Six queries (users, revenue, metrics(from,to), predictions(metric,horizon), anomalies(metric), insights(metric)) with structured Trend, ForecastPoint, and Insight types.
• AI insights service. OpenAI receives the already-computed numbers and is told never to recompute them. Process-lifetime cache.
• Dashboard. Metric switcher, four KPIs, Recharts time-series with dotted forecast and anomaly dots, AI insights panel, embedded GraphQL playground.
• "AI explains, never calculates", enforced by types. The engine module exports plain functions over number[]. The AI service takes the engine's Prediction object and turns it into a user-prompt block. There is no codepath where the LLM produces a number the dashboard renders.
• Server-side GraphQL execution via graphql.execute(); the HTTP endpoint stays for the playground and external clients.

Tradeoffs

• Linear regression and z-score, not ARIMA or Prophet. Interpretable, dependency-free, and good enough for the demo's signal shape. The engine boundary is plain (number[]) => Prediction, so a heavier model swaps in without touching the AI service or the GraphQL schema.
• LLM never produces numbers, enforced by types. Costs nothing in flexibility because arithmetic was the wrong job for the model in the first place; the constraint pays off on every refactor that might otherwise tempt a "quick" math call.
• Seeded synthetic dataset. Byte-identical across runs, which is great for code review and screenshots and useless as a load or anomaly-detection stress test. Bring real data for that conversation.

Live demo →