AI Analytics Dashboard.
Predictions that drive decisions.

The challenge

A fintech startup's operations team was making inventory and staffing decisions based on spreadsheets and gut feel. Their existing BI tool showed historical data but couldn't forecast trends. By the time they spotted a pattern, it was too late to act. They needed a system that could ingest live data streams, detect anomalies in real time, and project key metrics 30-90 days into the future with actionable confidence intervals.

The approach

I designed a two-layer architecture: a Python/FastAPI backend handling data ingestion, feature engineering, and ML inference, paired with a React frontend using D3.js for interactive, real-time data visualizations. The ML pipeline uses an ensemble of LSTM and Prophet models, automatically selecting the best performer per metric.

Key technical decisions:

• Ensemble forecasting — LSTM networks for non-linear patterns combined with Prophet for seasonality and trend decomposition, with automatic model selection based on backtesting MAPE
• Streaming data pipeline — Apache Kafka ingests live data sources, with Flink processing for real-time feature computation and anomaly detection
• Interactive D3.js visualizations — zoomable time-series charts with confidence intervals, drill-down capabilities, and annotation support for team collaboration
• Explainable AI layer — SHAP values surface the top 5 factors driving each prediction, building trust with non-technical stakeholders
• Automated retraining — models retrain weekly on fresh data via Airflow DAGs, with drift detection triggering emergency retrains when distribution shifts are detected

The outcome

• 91% forecast accuracy (MAPE) across all tracked metrics
• 2M+ data points processed daily with sub-200ms query response times
• $180K annual savings from optimized inventory decisions
• Anomaly detection catches issues 4 hours earlier than manual monitoring
• Executive dashboard adoption: 100% of C-suite uses it in weekly reviews
• 30-day, 60-day, and 90-day forecasts with calibrated confidence intervals

Technical highlights

The frontend renders 50,000+ data points smoothly using canvas-based D3 rendering with WebWorker-powered data aggregation. Time-series charts support sub-second zoom and pan across 2 years of historical data. The dashboard loads in 1.8 seconds thanks to server-side data pre-aggregation and progressive chart hydration.

The ML serving layer uses TensorFlow Serving behind an Nginx reverse proxy, handling 500 prediction requests per second with p99 latency of 45ms. Model versioning and A/B testing are managed through MLflow, enabling safe rollouts of improved models without downtime.