I build things at the intersection of AI infrastructure and language. Usually things that don't exist yet.
Lately I've been obsessed with how much a single person can ship when they stop context-switching and just let the tools rip. Claude Code, Codex, the whole agentic coding stack. RunAgent was largely built this way. What used to take a team now fits in one person's flow.
Currently shipping RunAgent, a serverless runtime for AI agents using Firecracker microVMs. Sub-second cold starts, full isolation, runs LangGraph/CrewAI/LlamaIndex agents like they're serverless functions.
Before that: built the first Bengali LLM (7B params, ~35B tokens) because it didn't exist. Also spent a year at Composio generating the tool layer that most AI agents quietly depend on.
RunAgent Serverless agent runtime on Firecracker microVMs |
RunAgent Genie 13-level prompt injection game, each run in an isolated VM |
RunAgent OSS Open-source Agent deployment runtime |
runagent-pulse Scheduling Calender for AI Agents |
superbrowser The browser your agent won't get blocked on |
at Intelsense AI I was CTO. built these because Bengali had no real AI infrastructure.
UniSense Multi-agent sales and support |
SenseForm AI-driven conversational forms |
Sense.ai Bengali LLM chat platform, Ekush LLM |
research side: Google Scholar · 5 papers · 65+ citations · GSoC 2023 Intel OpenVINO · IEEE VIP Cup 🥇 2021
python typescript rust go · pytorch huggingface langgraph · firecracker docker k8s · claude code codex