01 / DISTRIBUTED CONTROL PLANES
Making GPU clusters resilient and elastic
As a founding HyperPod engineer, designed desired-state reconciliation, continuous scaling, and resilient node operations for large foundation-model training clusters.
Senior ML Engineer / AI systems
Building AI systems at scale.
I'm Monidipa Chakraborty. I build RL training, evaluation, and data systems for AI coding agents—drawing on over a decade of experience engineering reliable distributed platforms and large GPU clusters.
01 / About
Systems thinking.
Research curiosity.
My work sits at the intersection of applied machine learning and the engineering needed to make it reliable, measurable, and scalable.
I'm a Senior ML Engineer at Amazon Kiro, building the training, evaluation, and data infrastructure behind AI coding agents. My work covers RL rollout quality and reward-hacking audits, coding-environment validation, agent-trajectory analysis, and distributed GPU infrastructure.
I've also worked hands-on with GRPO training using PyTorch, SLIME, Megatron, and SGLang—connecting the training loop with the operational systems that support it.
Previously, I was a founding engineer on AWS SageMaker HyperPod's control plane, designing continuous cluster scaling, desired-state reconciliation, and resiliency systems for foundation-model training clusters with thousands of accelerators.
I'm especially interested in reinforcement learning, coding agents, distributed training, and high-performance ML infrastructure—and in sharing practical lessons from building real systems.
02 / Selected work
Engineering at
the edge of scale.
Three areas where I've gone deep—from GPU cluster control planes to the quality of the feedback used to train coding agents.
02 / RL EVALUATION
Improving the signal in coding-agent RL
Built pipelines for reward-hacking audits, judge calibration, environment validation, and benchmark integrity—helping distinguish real progress from misleading rewards.
03 / TRAINING INFRASTRUCTURE
Operating distributed ML platforms
Built multi-region GPU training infrastructure and worked across scheduling, networking, petabyte-scale shared storage, observability, and recovery for ML workloads.
03 / Experience
From cloud systems
to AI systems.
A career spent building infrastructure that must work reliably at scale—and increasingly applying that discipline to AI development.
2026 — Present
AMAZON / KIRO
Senior ML Engineer
RL evaluation, coding-agent data systems, hands-on training experiments, and distributed GPU platforms for AI coding agents.
2023 — 2026
AWS / SAGEMAKER HYPERPOD
Senior Software Engineer · Founding Control-Plane Engineer
Architected scaling, reconciliation, and resilience systems for large foundation-model training clusters; led major launches and cross-team engineering work.
2014 — 2023
AMAZON & AWS
Software Engineering
Worked on SageMaker Processing and Batch Transform, Just Walk Out systems, and Prime Video—spanning cloud infrastructure, distributed services, and consumer products.
04 / Field notes
What I'm exploring.
A home for technical writing, reproducible experiments, and practical notes on the systems behind AI progress.
RL training systems01
Inside a coding-agent RL training pipeline
Rollouts, reward signals, training, and where systems bottlenecks emerge.
Planned topic / not yet published
GPU infrastructure02
Scaling GPU clusters without blocking operations
Desired-state controllers, reconciliation, and failure-aware orchestration.
Planned topic / not yet published
Evaluation quality03
When reward signals mislead coding agents
Reward hacking, environment integrity, and measuring actual progress.
Planned topic / not yet published
These are article ideas I'm developing, not links to published work.
05 / Connect
Let's talk about hard problems.
Especially RL systems, coding agents, distributed training, and the infrastructure that makes AI work at scale.