I build multi-agent systems that turn weeks of ML work into hours.
I’m Arindam Chowdhury, an Applied Scientist at Amazon working on agentic AI systems, LLM post-training, and production ML at scale. I design multi-agent systems that turn research into shipped, real-world impact.
What I work on
Systems that put large models to work, reliably, in production.
My current focus is agentic AI: orchestrating multiple LLM-driven agents to automate complex machine-learning workflows, from planning through execution and validation. Underneath sits a decade of applied-ML depth in deep learning, graph neural networks, and large-scale training.
Agentic & multi-agent systems
Planner–executor architectures, tool use, and agent harnesses for automating multi-step ML work.
LLMs & post-training
Fine-tuning and adaptation, RAG pipelines, synthetic data, and rigorous evaluation design.
Foundation & graph models
Graph foundation models and global-attention architectures, trained at supercomputer scale.
Production ML at scale
Real-time, low-latency inference pipelines serving worldwide traffic in production.
Selected projects
Multi-agent orchestration for ML automation
Planner–executor agents · production code generation
A multi-agent orchestrator that compresses multi-week ML pipeline launches into hours. It parses dense requirement documents into knowledge graphs for planning, grounds decisions in persistent cross-launch memory and an ML knowledge base, and drives a coding agent that generates production code with multi-step validation.
Agentic “co-scientist” for ML teams
Orchestrator + multi-agent executor · governance layer
An org-scale agentic system that automates routine scientist operations across experiment execution and communication, serving multiple ML, engineering, and PM teams. A governance layer steers agent behavior to each scientist’s style and prevents leakage of sensitive information.
RIPE — PU learning for early risk detection
Decomposed propensity · autoresearch loop
A positive-unlabeled learning framework for proactive, early detection that decomposes label propensity across channels, with an autoresearch loop optimizing propensity estimation against downstream AUC. Validated through production A/B testing.
Graph foundation models on Frontier
Global-attention GNN · leadership-class HPC
A systematic evaluation of fine-tuning strategies for a scalable global-attention graph foundation model (HydraGNN), run on the ORNL Frontier supercomputer. The post-training methodology transfers directly to LLM fine-tuning.
Research & patents
Selected publications
Granted patents
Full list on Google Scholar.
Background
PhD advised by Dr. Santiago Segarra, focused on graph ML, deep reinforcement learning, and distributed learning. Earlier degrees from IIT Guwahati (M.Tech, Signal Processing) and NIT Durgapur (B.Tech, ECE).
Contact
Open to conversations about applied AI and agentic systems.
The fastest way to reach me is email, or connect on LinkedIn.