Deepan Anbarasan

Restructured a production LangGraph + FastMCP agent architecture for multi-turn conversational workflows by splitting a monolithic planner into single-responsibility nodes with explicit state management, loop detection, and token-pressure context summarization.

Built and shipped Credence, a time series foundation model ranked top-5 on GIFT-Eval, from research through a production FastAPI API. Designed a governed Agent-as-a-Service with MCP endpoints and tiered RBAC/JWT access, plus sub-100ms inference, drift-triggered retraining, and anomaly detection.

Architected an end-to-end automatic passenger counting system spanning ESP32 firmware, MQTT, Python microservices, AWS, and ML/DL classification. Achieved 82% real-world accuracy with a Wi‑Fi probe-request method at 50%+ lower hardware cost than commercial optical alternatives.

Conducted PhD research on Wi‑Fi sensing and machine learning for automated passenger counting, benchmarking CNN, LSTM, and gradient boosting on noisy sensor data. Delivered systematic ablation results that guided production architecture later commercialised via MobyForAll.

Improved a labour forecasting and management tool deployed across 1,000+ stores, reducing loading time by 30% and delivering an estimated $650K in annual savings. Enhanced remuneration management, reducing supervisor processing time by 10%.

Developed backend modules and provided production support for a workforce management platform for a major retail chain, using Java, Spring, and Oracle SQL. Contributed to ongoing reliability and operational maintenance of the system.

Grade: cum laude

PhD in Urban and Regional Development (cum laude) focused on ICT in public transport using IoT and machine learning for automatic passenger counting. Conducted research informing production architectures for passenger counting systems.

MSc in ICT for Smart Societies with a thesis on IoT and data analytics for mobility patterns of public transport users.

BEng in Electronics and Communications Engineering.