I'm Monidipa Chakraborty. I build evaluation and data systems for coding agents, trustworthy execution environments, and the distributed GPU infrastructure behind modern AI—drawing on 12 years of engineering at Amazon.
I work where coding-agent behavior, rigorous evaluation, and infrastructure meet—making complex AI systems more trustworthy and scalable.
I'm a Senior ML Engineer at Amazon Kiro, building production agent-trajectory data pipelines, LLM-assisted labeling, privacy-preserving task taxonomies, and benchmarks used to understand how coding agents perform.
I also work on coding-environment creation and validation, reward-integrity audits, and LLM-judge calibration—catching faulty environments, misleading success signals, and evaluation leakage before they distort results.
Previously, I was a Founding Engineer at AWS SageMaker HyperPod, where I helped build the systems for continuous cluster scaling, desired-state reconciliation, and resilient operations across large GPU fleets.
Across these roles, the common thread is turning hard operational and measurement problems into reliable, well-instrumented systems. I enjoy sharing what I learn through writing and open-source experiments.
From agent-data pipelines and coding-environment quality to GPU infrastructure that operates at cloud scale.
Built production trajectory labeling, privacy-aware task taxonomies, and evaluation datasets for coding agents. Combined LLM extraction, embedding-based clustering, and controlled prompt experiments.
Designed environment creation and pre/post-run validation, audited reward integrity, and uncovered solution leakage that could inflate coding-agent benchmark results.
As a Founding Engineer at SageMaker HyperPod, helped build cloud-scale cluster workflows, continuous scaling, and resilience mechanisms; later worked across multi-region GPU scheduling, networking, and shared storage.
Twelve years of engineering at Amazon, spanning distributed platforms, foundation-model infrastructure, and the systems used to measure AI coding agents.
Agent data and evaluation pipelines, privacy-aware task taxonomies, coding-environment quality, reward-integrity audits, and ML platform infrastructure.
Built cluster lifecycle workflows, continuous non-blocking scaling, reconciliation, multi-tenant isolation, and resilience for foundation-model GPU infrastructure.
Built and operated systems across SageMaker Processing and Batch Transform, Just Walk Out, and Prime Video.
Selected public repositories—from reproducible coding-agent benchmarks to hands-on model-building exercises and developer tooling.
An exploratory long-horizon coding benchmark using Harbor and OpenSandbox, including network-isolated verification, pinned environments, and a task-design rubric.
A hands-on PyTorch learning project inspired by Karpathy's lectures, covering character-level language models, embeddings, batch normalization, and manual gradient checks.
Reproducible terminal, editor, and shell configurations across Zsh, Helix, Neovim, Ghostty, and other tools, managed with GNU Stow.
Occasional writing on machine learning, AI infrastructure, and what it takes to turn technical ideas into working systems.
A short perspective on the compute race: why access to GPUs, clusters, and reliable infrastructure shapes the pace of AI progress.
An introduction to supervised, unsupervised, and reinforcement learning, plus the core steps of building an ML system.
Especially agent evaluation, reproducible coding environments, resilient GPU platforms, and making complex systems work in practice.