Manas Choudhary

Running AI agents on your own GPU means wiring together model serving, memory, safety checks and code execution by hand. Then you rebuild all of it for the next agent.

I built AgentOS, a control layer for self-hosted agents. An agent changes one URL, and each extra (memory, safety, sandbox) is switched on from a manifest file.

It carried 383,254 requests from 50 to 3,200 users with 0% errors, and 50 crash tests lost no audit events. One target I missed: realtime p99 is 202.1 ms against a 200 ms goal.

Nobody knows which AI agent called which model, what it cost, or whether it leaked personal data. Fixing that usually means adding a new library to every agent.

I built Govrix Scout, an open-source proxy in Rust. Agents change one web address, and every request crossing it is identified, checked and recorded.

It handled 383,254 requests from 50 to 3,200 users with 0% errors, and the open-source repo has 70+ GitHub stars. It is also the front door of AgentOS, where I hardened it further.

Teams building a chatbot over their own documents have to pick how to cut the text up, how to search it and how to re-rank results. Most people guess, and nobody can tell if the guess was good.

ChunkScope looks at your documents first, recommends a setup with a confidence score and a reason for every option it rejected, then lets you build, run and grade it on a drag-and-drop canvas. I built it with Asmit Shaw.

The scan reads 20 signals from the text without calling an AI model, the guide covers 39 techniques, and the finished pipeline exports as plain Python with no LangChain. Its backend now runs inside AgentOS as MemoryMesh.

Standard AI benchmarks are saturating, and some questions probably leaked into training data, so scores say less and less about how models actually think.

With Prashil Thul, I built a head-to-head game where two LLMs place Conway's Game of Life patterns, a deterministic engine runs the fight, and a tournament ranks the models with Bradley-Terry ratings.

Nine frontier models played 72 matches for about $66. The top-rated model spent $1.78 in total, but the top three ratings overlap, so two games per pair could not separate them.

Turning a Figma design into working code is slow manual work, and AI tools that try it often claim success while the page looks wrong.

Aura is an agent that reads a Figma design and writes React and TypeScript. It then renders the result in a real browser, scores it against the design and fixes it until it passes. I built it as an R&D intern on Samsung PRISM.

The second version replaced a heavy multi-agent setup with one supervisor agent. The config requires a 95% match to pass and allows up to 10 fix rounds. The first version, Aura1, was abandoned.

When production breaks at 3am, an engineer can spend 40 minutes digging through logs to find out why a service keeps crashing.

I built a team of AI agents that watches a Kubernetes cluster, finds the fault, diagnoses it and proposes a fix. A safety gate decides whether the fix runs on its own or waits for a human.

The gate auto-runs a fix only when three checks pass, and everything else pauses for Slack approval. The aim was under 90 seconds per incident, but I did not measure that, so it is a goal.