MarketLens
LIVE • PRODUCTIONFully Automated AI Trading Intelligence
A production SaaS platform built on Google Cloud Run. It orchestrates a 7-model neural ensemble — Transformer, CNN-LSTM-Attention, TCN, N-BEATS, LSTM-GRU, Enhanced Informer and XGBoost — across a zero-touch automation pipeline. Real-time data ingestion feeds automated feature extraction of 50+ indicators, flows through ensemble inference and a 5-check quality gate, and delivers results via Telegram and API. Signals cover 19 instruments spanning crypto, forex, commodities and indices, across three timeframes with sub-second inference.
Self-Healing MLOps
Scheduled GPU retraining on Kaggle, GCS model versioning and webhook-driven hot-reload. Zero-downtime deployments with no redeploy.
SHAP Explainability
Per-prediction feature importance analysis for model transparency and compliance. Every output is interpretable and auditable.
Risk Controls
Fixed-fractional position sizing, volatility-scaled stops, graduated profit targets and max drawdown limits per asset class.
Automated Billing
Stripe integration with 3 subscription tiers, Firebase Auth, per-user Firestore data isolation. Event-driven, fully automated.
dbt Data Pipelines
dbt Core transformation layer with 22 automated tests across 1,718 predictions, 461 execution records, 252 drift logs, 242 equity snapshots.
RAG Chatbot
Gemini embeddings (768-dim), Firestore cosine KNN vector search, LLM generation over 29-document knowledge base for intelligent Q&A.
EnergyLens
LIVE • PRODUCTIONNordic Energy Market AI Forecasting
A live forecasting platform for Nordic power markets (DK1/DK2). Built with bitemporal data pipelines for point-in-time replay, a 6-model neural ensemble for 24-hour-ahead price prediction, and a 5-gate signal quality framework with SHAP explainability and forecast accuracy tracking. Ingests Nord Pool spot prices, Open-Meteo weather data and ENTSO-E generation data on an automated six-hourly schedule, with 568,681 generation records backfilled across DK1 and DK2 to train on a full year of wind, solar and thermal output.
Bitemporal Pipeline
valid_time and knowledge_time columns enable point-in-time historical replay. Forward-fill non-price features, temporal advancement, absolute-mode clamping.
SHAP Explainability
Grouped feature importance analysis (temporal, weather, price history) providing transparent, auditable rationale for each forecast cycle.
Accuracy Tracking
Automated engine comparing predictions against actuals with per-model MAE/RMSE breakdown, historical trends, and directional accuracy metrics.
Generation Features
568,681 ENTSO-E records across DK1/DK2 drive wind, solar and total-output features — lags, rolling means, volatility and price-per-MW ratios.
Zero-Touch Refresh
Cloud Scheduler triggers ingestion, feature build and forecast every six hours. Retrained models hot-reload from GCS without a redeploy.
Ensemble Weighting
Model outputs blended by inverse-MSE weighting derived from walk-forward cross-validation — no single point of model failure.
Honest Monitoring
12,000+ forecasts logged and scored against actuals across 24/48/72h horizons. Systematic bias in the tracker is what triggers retraining.
Lean Ops
React 18 + Vite dashboard on a FastAPI backend with 15+ REST endpoints, containerised on Cloud Run and running inside the GCP free tier.
Signal Quality Gate
Five gates — confidence band, model consensus, data freshness, forecast stability and volatility — block low-trust forecasts before publication.
