Staff Forward Deployed Engineer, Delta @ Google Cloud — Singapore
I ship agentic AI into production, and publish the research behind it. Multi-agent systems · AI safety · LLM evals · inference optimisation
Speaking / collaborating? → contact@subhadipmitra.com · subhadipmitra.com
- Verifiers & Post-Training: RLVR, reward hacking, and evaluation that holds up under optimization (rlvr-rubric-hacking)
- Agent Infrastructure: Governed MCP tool surfaces (api-mcp-compiler), loop engineering, human approval gates
- AI Interpretability & Safety: Activation probes, steering vectors, sandbagging detection, the runtime cost of interpretability
- LLM Inference: Triton kernels for MoE dispatch and weight-only quantized GEMM (triton-kernels)
- Distributed Systems: Formal verification, UPIR
- When a Good Reward and a Lousy Reward Look the Same in the Top Decile — reward hacking in RL, and why top-decile appearance misleads.
- Why "one loop to rule them all" fails in production AI agents — a new abstraction of agentic loops as effects.
- Reward models & verifiers
- Agent infrastructure & MCP tool surfaces
- Reward hacking mitigation
- Interp, steering, and inference optimization
- Agent evaluation & post-training (RLVR & cap-based credit assignment)