An event-driven AI logistics platform that pairs classical Operations Research (OR-Tools CVRPTW) with a five-agent LangGraph decision layer. The architectural thesis: LLMs decide when and how to invoke the solver — they never replace it. Mathematical optimisation is authoritative; agents handle anomaly detection, SLA risk scoring, and re-solve decisions around it.

• ›Five-agent LangGraph pipeline: Monitoring → Classification → SLA Risk → Rerouting → Dispatch, with conditional edge skipping ~70% of LLM calls on nominal ticks
• ›OR-Tools CVRPTW solver with 7 hard constraints (capacity, time windows, shift limits, depot routing) — provably optimal route plans
• ›Idempotent event processing via DuckDB anti-join — agents never re-process the same event regardless of run count

An autonomous virtual power plant (VPP) orchestrator that parses natural language policy documents into linear program constraints, forecasts solar energy using XGBoost, and generates mathematically optimal battery charge/discharge schedules via Google OR-Tools.

• ›Deterministic three-phase pipeline: Gemini Intent Parser (using response_schema) → Qdrant Policy Retrieval → XGBoost Solar Forecast → OR-Tools LP Solver
• ›Dynamic RAG-to-constraint mapping: parses regulatory SOPs at runtime to inject safety buffers (e.g. 30% for hospital reserve) without code changes
• ›Asynchronous worker architecture: Celery + Redis offloads blocking solver computations from FastAPI to handle high-concurrency request polling

A cloud-native decision intelligence platform that turns energy transition policies into optimized strategies for hybrid energy systems: gas, renewables, storage, and hydrogen.

• ›MILP optimization engine (Pyomo) minimizing cost, emissions, and unserved energy
• ›ML forecasting: LSTM + XGBoost + Prophet for demand and renewable prediction
• ›Monte Carlo simulation for uncertainty in demand, fuel prices, weather

A cloud-native pipeline that handles the full lifecycle of energy distribution data: simulation, streaming, processing, storage, transformation, and optimization.

• ›Real-time IoT simulation → Kafka → Spark → PostgreSQL + S3 pipeline
• ›Stage 1 LP optimizer: fair energy allocation across 5 zones by priority
• ›Stage 2 Transportation optimizer: routes power minimizing transmission loss

A full-stack system that turns life goals into mathematically optimized daily schedules using Mixed-Integer Linear Programming.

• ›MILP solver (PuLP + CBC) optimizing daily schedule across time, energy, and behavioral constraints
• ›Solves typical schedules in under 500ms
• ›JWT authentication, role-based access, async FastAPI backend

A decision intelligence system fusing XGBoost survival prediction with Mixed-Integer Programming to solve lifeboat allocation under ethical and capacity constraints.

• ›XGBoost classifier achieving 85% accuracy and 0.82 F1 score
• ›MIP optimizer allocating lifeboat seats by survival probability + ethical constraints
• ›Ethical constraints: ≥30% children, ≥50% women priority guarantees