Manas Choudhary

AgentOS

Run and govern AI agents on hardware you own

Every AI agent needs model serving, memory, safety and code execution, and teams rebuild them for each agent. AgentOS gives any LangChain, LangGraph or OpenAI-SDK agent all four through one URL change, each switched on from a manifest file.

• Carried 383,254 requests from 50 to 3,200 users with 0% errors
• Schedules many agents on one RTX 5090 with priority classes: 280 real preemptions in a 5-agent run
• Benchmarks picked one Postgres over Qdrant and Neo4j
• Realtime p99 missed its 200 ms target by about 2 ms, and the page says so

Govrix Scout

Open-source governance proxy for AI agents · 70+ stars

Nobody knows which agent called which model, what it cost, or whether it leaked data. Govrix is one door every agent request crosses: change one base URL and each agent gets a kill switch, PII checks and a SHA-256 hash-chained audit log, with no SDK.

• Audit events go through a bounded queue, so the database never sits on the request path
• Fails open: if its own database breaks, agents keep working and a warning is logged
• 726 Rust tests across a 9-crate workspace

ChunkScope

Visual RAG pipeline builder · built with Asmit Shaw

Choosing how to chunk, retrieve and rerank documents for RAG is guesswork. ChunkScope scans the corpus first, recommends a technique stack with a confidence score and a reason for every rejected option, then lets you build, run and grade the pipeline on a canvas and export it as plain Python.

• Corpus scan: 20 signals in under 2 seconds, with no LLM call
• Confidence scores calibrated from published benchmarks, 48 techniques in total
• Exports pipeline code with no LangChain dependency; 162+ tests

Google DeepMind Concordia

Open-source contributor · 1.7k stars

Memory lookups in DeepMind's agent-simulation library slowed down sharply as memories grew. 7 of my changes are merged, including a memory-store rewrite from quadratic to linear time: 10 s down to 0.1 s for 1,000 records.

Game of Death

LLM benchmark · built with Prashil Thul

Two LLMs play competitive Conway's Game of Life. Each round a model places patterns with tool calls, 160 generations run, and enemy cells that reach its Heart cost HP. The engine is deterministic, so the only variance is the model. We ran 9 frontier models in a seat-swapped round-robin for about $66.

• 72 matches, Bradley-Terry ratings with 95% confidence intervals
• The top-rated model cost $1.78 for the whole tournament; the runner-up cost $22.58
• With 1 round every game drew; with 5 rounds, 43 of 71 games had a winner

Aura

Figma-to-React agent · Samsung PRISM

Converts Figma designs into production React and TypeScript, then refuses to trust its own success: it renders the result with Playwright, scores the screenshot against the Figma export, and retries until the render clears a threshold or hits a bounded iteration cap.

K8sWhisperer

Kubernetes incident responder

Auto-executes a Kubernetes remediation only when all three hold: model confidence above 0.8, blast radius low, and the action absent from a destructive-action list. Everything else suspends the graph and waits for Slack approval.