I'm Vinayak Dubey, a software engineer in Bengaluru. I work at Oracle on AI agent systems, and I have about two years of professional experience. Most of my time goes into making agents that do real work behave reliably once they leave the demo.
My day-to-day is Python backend services, agentic pipelines built on LangGraph, and tooling that exposes internal systems to models through the Model Context Protocol (MCP). Projects I have worked on include natural-language-to-SQL services over analytics data, MCP servers that let models call internal APIs, and agent-driven dashboards. I have also mentored interns at Oracle.
Before this I worked across the full stack with React, TypeScript, FastAPI and LangChain, shipping on Docker, AWS and Azure. I hold a B.E. in Information Science from BMS College of Engineering.
I created Kestrion, an open-source framework for building production AI agents. It is published on PyPI. The idea is simple: when an agent crashes halfway through a task, it should resume where it stopped instead of starting over. That means an event log as the source of truth, approval steps before risky actions, and safe handling of parallel tool calls.
I am early in my career. I write down what I learn as notes rather than final answers, and I try to say plainly what I know and what I am still working out.
What I work on
- Agents in production. Agentic pipelines, retries, timeouts, state and observability.
- Durable execution. Making long-running agents survive failures and restarts. This is what Kestrion is for.
- Tool integration. MCP servers and well-shaped tools that models can use correctly.
- Natural-language-to-SQL. Turning questions into safe, validated queries.
- Backend engineering. Python, FastAPI and services that run without babysitting.
Tools I use
- Languages. Python, TypeScript, JavaScript, SQL.
- Frameworks. LangGraph, LangChain, FastAPI, React, Next.js.
- Infrastructure. Docker, AWS, Azure, GitHub Actions.