About Me
I am an AI Scientist at Nantum AI, where I develop AI-driven solutions for building systems—including HVAC fault detection, air-side anomaly detection, LLM-based diagnostic recommendations, adaptive modeling for static pressure, Model Predictive Control, and Reinforcement Learning for air-side optimization—all deployed in AWS for real-world clients.
My work sits at the intersection of two core research pillars. The first is AI-based fault detection and prognostics across diverse physical systems: building HVAC infrastructure, industrial rotating machinery (bearings), and power distribution networks. I apply time-series analysis, Topological Data Analysis (TDA), Explainable AI (Grad-CAM), and survival analysis to detect anomalies, classify faults, and predict failures—enabling trustworthy, human-interpretable decision support in safety-critical environments.
The second pillar is intelligent planning, scheduling, and operations research for complex multi-agent systems. My Ph.D. (ADAMS Lab, University at Buffalo, advised by Dr. Souma Chowdhury) produced Graph Reinforcement Learning frameworks for Multi-Robot Task Allocation, Capacitated Vehicle Routing, Urban Air Mobility fleet scheduling, and power network reconfiguration—achieving real-time, scalable decisions under uncertainty that outperform or match state-of-the-art solvers at a fraction of the computation cost.
I also spent two years as a Software Engineer at Wayfair (Big Data, Apache Storm, AKKA, .Net). Outside of work, I enjoy reading, working out, hiking, piano, soccer, and astronomy.
Expertise
End-to-end AI pipelines for detecting, classifying, and predicting failures across physical systems—from building HVAC units to industrial rotating machinery.
Scalable, real-time decision-making for complex combinatorial problems in robotics, logistics, and infrastructure—using Graph RL to match or beat classical solvers.
Deep expertise in modern ML architectures for sequential decision-making, state estimation, and trustworthy AI across safety-critical domains.
Experience
Present
- Fault detection & diagnostics: AI-based HVAC fault detection and air-side anomaly detection from time-series sensor data; LLM-powered diagnostic recommendations deployed in AWS for client buildings
- Planning & optimization: Adaptive modeling for static pressure set-point scheduling; Model Predictive Control and Reinforcement Learning for real-time air-side energy optimization
Jul 2025
- Led Percev's Analytics Team; survival analysis on bearings for failure prediction and defect classification
- Architected CI/CD framework for analytics; Random Survival Forests with >75% accuracy
- Commercialized prediction model as a new product; built customer-facing data visualization UI
Jan 2025
- Fault classification & XAI: Explainable CNN (Grad-CAM) for bearing fault classification; lightweight trustworthy prediction framework with up to 22% accuracy gain for high-confidence samples
- State estimation: Probabilistic high-rate dynamic system state estimation using TDA and ML; avg. compute time <50ms, outperforming state-of-the-art methods
- Failure prediction: Survival analysis on rotating machinery; Random Survival Forests with >75% accuracy; commercialized as a product for Grace Technologies
Jan 2024
- Graph RL for MRTA, CVRP, UAM fleet scheduling, and power network reconfiguration
- Funded by ONR AI/ML (N00014-21-1-2530) and NSF CAREER (CMMI 204802)
Dec 2019
- Notifications infrastructure: email & push for all Wayfair customers; personalized content engine
- Supply chain routing (C# / .Net) for small and large parcel delivery across US and Canada
- Built load-testing frameworks; maintained apps in AKKA, Apache Storm, Dropwizard, PHP
Education
Selected Publications
View full publication list on Google Scholar →
Projects
Capacitated Vehicle Routing Problem (CVRP)
Scalable Graph RL framework for CVRP with up to 50% cost reduction and 20× faster computation vs. meta-heuristic methods, without retraining for larger instances.
Multi-Robot Task Allocation (MRTA)
GNN + RL framework for MRTA with up to 6% higher task completion and 10× faster computation than meta-heuristics, scalable to larger problems without retraining.
MRTA – Collective Transport
GNN policy augmented with Topological Data Analysis for multi-robot collective transport. Up to 10% higher task completion vs. other learning-based methods; decision time in milliseconds.
Urban Air Mobility Fleet Scheduling
Learning-based UAM fleet scheduling with >12% revenue improvement over traditional RL, considering demand, passenger fare, and electricity pricing. Decision time in milliseconds.
Power Network Reconfiguration
Graph RL for optimal network reconfiguration and power flow, achieving comparable performance to conventional methods with >10× faster computation.
UAM Vertiport Management
Graph RL for real-time vertiport take-off and landing scheduling under environmental uncertainties.