Sameer Chaturvedi

Prime Video Personalization and Discovery, New York

• Core developer of Son of Anton, the team's agentic AI oncall platform: 7 Rust tools and 17 agent skills that give LLM agents live access to production systems, run in Claude Code and Kiro. It became the team's standard way to investigate production issues.
• Shipped one of the first AWS Bedrock AgentCore agents with per-user OAuth on-behalf-of delegation: platform GA to a production runtime in 8 days, now live in Slack answering as each requester.
• Worked across two generations of a worldwide subscription-channel recommender: a contextual multi-armed bandit learning in 6 regions, and its replacement, a sequence-aware collaborative-filtering model served through approximate-nearest-neighbor retrieval, whose 8-market A/B launch I owned.
• Root-caused months of silent recommendation degradation in one region to a stale upstream input, shipped six iterations of data-freshness safeguards, recovered the fleet across six regions, and built an automated health monitor (PySpark, Airflow on EMR on EKS, 80 tests).
• Argued for a final A/B test before a planned carousel retirement. It showed significant subscription lift (+134K annualized), reversed the decision, and led to worldwide expansion.
• Delivered a cross-org Fire TV home-screen filtering cap, caught four silent-failure modes before ship, and built and ran the partner's production load tests, cutting the required load about 3x.
• Owned annual infrastructure cost planning for two rounds (about 7.5% year-over-year reduction against a 5% target) and tier-1 oncall: 11 primary weeks and about 20 Sev2 incidents.
• Won the 2026 Prime Video hackathon with an automated NBA highlight row: live stats and play-by-play, PyTorch neural style transfer, OpenCV, ComfyUI, and stat-driven title cards.

Disruption Studies Group -- DisruptionPy

• Architecting DisruptionPy: a framework for analysing plasma-fusion data across five tokamak reactors to predict and avoid disruptions using anomaly models.
• Revamped the backend to decouple data access from specific implementations: pluggable backends (MDSplus, Xarray/Zarr) behind abstract interfaces for testability and multi-machine scalability. Validated bit-identical output against the reference dataset before merge.
• Root-caused and fixed a family of multiprocessing start-method bugs (fork vs spawn vs forkserver) that silently dropped every MAST physics variable on macOS and Python 3.14.
• Moved framework defaults into layered configuration (TOML, environment, explicit arguments) so each run records the settings it actually used.
• First author on an APS Division of Plasma Physics 2026 poster on maintainable, extensible architecture for multi-machine disruption studies.

Live Gameplay Team

• Short engagement in summer 2025, scoped to one demo for a tournament.
• LLM chat over live tournament stats, with realtime graphs and dashboards.
• Designed the pipeline that converts semantic text to SQL, queries the stats DB, processes data, and handles edge cases.

M365 Security Detections

• Petabyte-scale big-data platform using ML to detect cyber-attacks in near real time across all of M365.
• Lead engineer for Microsoft's first mail-anomaly detection (response to the nation-state incident where hackers accessed executive and federal emails).
• Lead engineer on a GPT-based security-detection assistant that analyses machine and identity activity for attack patterns.
• Designed the production workspace that enables data-science and LLM-backed detections in an agile PySpark environment with CI/CD and realtime job monitoring.
• Expedited detection programming by over 75% on distributed systems; auto-failover geo-replicated BCDR architecture.

Computer Vision Team

• Lead engineer on a deep-learning + iOS app for real-time object detection and instance segmentation.
• Improved accuracy with dropout, batch normalisation, and training on custom synthetic data.
• Converted models to run on-device via LibTorch; wrote the Objective-C++ bridge, hand-rolled NMS, and the full AVFoundation capture pipeline. CoreMotion integration plotted phone-path trajectory.
• Productionised the library and ETL pipeline with custom PyTorch DataLoaders and dataset mappers.

OpenShift -- Container Runtimes

• Upstream contributions to CRI-O (Kubernetes container runtime interface) and Libpod (container and pod manager).
• Unblocked engineers on image-volume write issues inside containers, fixed container-hook monitoring and execution, and added CLI env-var flag passthrough.