The layer that decides what an agent is allowed to do, and proves what it did. Currently on Atlan's AI-native SDLC platform, where the autonomous code-review agent I own has run 14,500+ reviews across 300+ repositories.
Self-hosted control plane for AI agents
Runs any agent behind one OpenAI-compatible endpoint that carries its spend budget, retrieval namespace, sandbox and audit trail with it, so governance travels with the agent instead of being bolted on per deployment. Storage is six databases rather than one general-purpose engine, because time series, vector search, hot state and graph traversal have genuinely different read shapes.
Governance proxy for AI agents · 72 stars
One environment variable puts it in front of any agent's LLM calls, so a team gets per-agent spend budgets, PII and jailbreak scanning and a tamper-evident audit trail without touching agent code. It drops audit events under queue pressure rather than stalling live proxy traffic, and gives streaming responses a separate unbounded mirror so observing a stream never slows the stream it is watching.
Open-source contributor · 1,658 stars
Landed 4 merged fixes, then replaced the full DataFrame copy running on every memory write with a deferred flush, taking agent memory insertion from quadratic to amortized linear.
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.
Visual RAG pipeline builder
Build two retrieval pipelines side by side and score them on the same document, so a chunking or reranking choice is settled by a RAGAS number instead of an argument. Supports 8 chunking strategies, hybrid retrieval (BM25, HyDE, MMR), cross-encoder reranking, and Python export of whichever pipeline won.
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.
Four agents over one syntax tree
A VS Code extension that reads a codebase four different ways - refactor, architecture, dependency, teaching - off a single AST pass, with multi-dimensional scoring to pick which chunks each agent sees. The project that got him the Atlan internship.
Career matching on a scored profile
Six agents build a candidate profile from Big Five and RIASEC assessments, then score career fit against it rather than keyword-matching a resume. React 19 and TypeScript front end, FastAPI back end.
TECH UNIVERSE
Python
JavaScriptTypeScript
C++
HTML5
CSS3ReactNext.js
Tailwind
Three.js
FramerNode.js
Express
FastAPI
Flask
GraphQL
Playwright
TensorFlow
KerasLangChainCrewAIRAG
LLMLangGraph
MCP
MongoDB
Postgres
Firebase
Neo4j
Redis
AWS
GCPDocker
GitGitHub
VercelReact Native
ExpoFive teams, one thing in common: agents that had to work in production, not in a demo.
Atlan
Feb 2026 - Present | Singapore HQ · Remote
Builds the execution and quality layers of Atlan's AI-native SDLC platform, used by 16+ engineering teams to put their own agents into production. Owns the autonomous code-review agent, which has run 14,500+ reviews across 300+ repositories, each inside its own sandbox with 23 security skills, filing critical findings as tracked tickets with no human in the loop. Fix pull requests it opens stay drafts behind a repository allowlist, highest severity first, and merging one is how a human accepts it. Runs the sandbox execution layer end to end: 100 to 150 isolated environments a day, session ownership settled in Postgres advisory locks and compare-and-swap rather than in racing pods.
Skills: LangGraph, FastAPI, Postgres, Kubernetes, Cloudflare Durable Objects