projects

Durable AgentAutonomous coding agent

Turns a GitHub issue into a pull request: plans the change, writes the code, runs tests in an isolated Linux sandbox and attaches the results. Memory and files persist between runs, and cost per task is down ~50× to $0.014.

TypeScriptCloudflare WorkersDurable ObjectsSQLiteSSE

live source

PassportPortable auth for AI agents

One login for every AI agent you run. Connect GitHub, Slack or a database once, approve an agent's request, and it gets a short-lived scoped pass you can revoke in one click. Ships a CLI, an MCP server and 90 providers.

TypeScriptExpressMongoDBReactMCP

live source

TrainOpsFine-tuning & inference platform

Fine-tune open-source models and robotics policies from the browser, then call them through one OpenAI-compatible API. Runs on serverless GPUs with queueing, checkpoint/resume, retries and live training logs.

PythonFastAPIPostgreSQLModalPyTorchLoRA

live source

VideoFlowVideo collaboration for YouTube teams

Editors upload drafts, reviewers comment on exact timestamps, and creators approve and publish to YouTube without sharing the channel password. Includes AI thumbnail generation.

Next.jsNode.jsMongoDBAWS S3OpenAI

live source

work

Accenture

Software Engineer · Sep 2024 to present

Backend & Applied AI · internal AI agent platform

Agent workflows in production. Shipped multi-step LLM agents on LangGraph and CrewAI that took over manual review and structured data extraction.
Tool runtime. Designed the shared tool layer every agent calls: typed schemas, scoped permissions and predictable errors, so new tools and agents plug in without touching the runtime.
Retrieval. Built RAG over embeddings and vector search, so agents and users find answers by meaning, not exact keywords.
Quality and observability. Built an eval harness that catches regressions when a prompt, model or tool changes, plus per-run traces of reasoning, tool calls and token cost.
Platform scale. Own the async Python and Node services underneath, serving 10M+ requests a month. Cut API latency by 30% with Redis caching, query rewrites and background workers.

also worked with folks at