Sarthak Singh

I’m an AI Engineer who builds reliable multi-agent systems end-to-end, with a focus on dependable execution and context stability. I developed a custom agent harness engine with per-agent sandboxed execution, tool-call routing, and output parsing to complete tasks across heterogeneous agent types without shared state leakage.

In my recent work, I designed an event-driven agent runtime that captures user interaction patterns and session history to adapt behavior across multi-turn conversations. I also implemented a 3-tier context management system (short-term, working, long-term memory), reducing context-related failures by ~40%.

Prior to that, I strengthened my ML and production foundations through infrastructure and backend roles. As an ML Infrastructure Engineer intern, I refactored a monolith to CQRS/Event Sourcing with Apache Kafka (improving throughput ~4x), implemented GPUDirect Storage to cut diffusion cold-start from 45s to 3s, and used CUDA streams to boost GPU utilization to ~85% and throughput ~6x on Kubernetes with Triton Inference Server.

I’ve also shipped backend systems on AWS with strong observability and automation. I built event-driven FastAPI services with RAG using LangChain, migrated a React monolith to Next.js with Turborepo (cutting build time ~60%), and developed a robotics portfolio in ROS/ROS2 with OpenVLA/SAM2 and Jetson-based vision-language-action—achieving 78% task success on pick-and-place and 94% collision-free trajectories in simulation.