About
Production multi-agent systems, and how agents should coordinate
Agent engineer in Bangalore. I ship multi-agent products in production and study multi-agent RL part-time at the University of Groningen.
I design and ship multi-agent systems in production: collaborative workflows, tool-using agents, fine-tuning, retrieval when it matters, and evals so teams can iterate without guessing.
I co-founded a voice-AI startup around multi-agent customer automation and hybrid data routing for agents. Before that I built a connected-vehicle data platform (10,000 vehicles in three months, later acquired). Earlier work spans industrial 6D pose estimation and LLM fine-tuning with QLoRA.
I am pursuing a PhD in multi-agent reinforcement learning at the University of Groningen, focused on strategic world models. I proposed SeqPPO, which showed roughly 3× better sampling efficiency than MAPPO, HATRPO, and HAPPO in our benchmarks.
Principles
- Use multiple agents when a single prompt chain stops scaling.
- If you cannot trace a multi-step run, you should not ship it.
- Pick the right layer: orchestration, fine-tuning, retrieval, or tools.
- When something breaks in prod, leave a runbook the team can reuse.
- Research on coordination informs how I design agent handoffs.
Education
- PhD, Multi-agent reinforcement learning, University of Groningen (2026–2030)
- MSc, Artificial Intelligence, University of Groningen (2020–2022). GPA 7.4/10
- B.E., Information Science and Engineering, Global Academy of Technology (2015–2019)
Recognition
- Winner, Maastricht WiDS Datathon, "Data Science Pioneers" (2022)
- Finalist, YALE-CBIT Hackathon (2022)
- Team lead, Ai4Good × European Space Agency (2021)
- Audience choice, AiMED:AiHack Covid (2021)
- University research collaboration; grant proposal secured $100K funding (2022)