Amazon Web Services (SageMaker AI)
Software Development Engineer (ML)
Seattle, WA
Hi, I'm Vedant
I'm a recent USC computer science grad working as a software engineer on SageMaker HyperPod in Seattle. Outside work I build agent evals that turn live failures into regression tests, post-training data tools, and LLM inference engines from scratch, with continuous batching, paged KV-caches and speculative decoding. I'm also extremely passionate about robotics having started my own FRC team and worked on robotics research at CMU.
Projects
Agent Regression FlywheelGitHub
Live agent failures become regression tests; promotions gated by a significance test.
HindsightOn request
Finds the decisive failing step in an agent run and turns it into SFT and DPO data.
CalibrateOn request
Scores post-training tasks with item response theory and flags the broken ones.
paged-speculative-engineGitHub
Continuous batching, paged KV-cache and speculative decoding, proven correct four ways.
Agent Ops ConsoleOn request
Reviewers flag agent mistakes; approved flags become regression tests.
inference-batching-labGitHub
KV-cached transformer plus three scheduling strategies, benchmarked. 2.2x throughput.
idempotent-payments-ledgerGitHub
Exactly-once charges and a balanced ledger under concurrency and retries.
agent-redteam-fuzzerGitHub
Prompt-injection fuzzer for tool-calling agents: 5 attack categories, 45 payloads.
agent-reliability-harnessGitHub
LLM proposes, deterministic guardrails decide, failures become regression tests.Experience
Jul 2026 – Present
Amazon (AGI Infrastructure)
Software Development Engineer Intern
Seattle, WA
Biorobotics Lab, Carnegie Mellon University Robotics Institute
Robotics Software Engineering Intern (Computer Vision)
Pittsburgh, PA
USC Viterbi School of Engineering
Teaching Assistant
Los Angeles, CA
Interests
Inference and ML systems
I like the part of ML where the hard problem is a systems one: scheduling, memory, correctness. I build engines from scratch so I can explain every line, then measure them honestly, including when the result is a negative one.
Robotics and FRC
FIRST Robotics Competition Dean's List Finalist. At CMU's Biorobotics Lab I built a ROS pipeline for a pipe-inspection robot, fusing 10,000+ laser-scan points with camera images into colorized 3D point clouds and simulating 1,000+ inspections.