Building data platforms and AI systems.
Google Cloud • Southeast Asia
- 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
- I Trained Three Models Against the Rubric I Published in January. Two Learned to Game It. - GRPO against my own RLVR rubric: two models gamed it, one did not
- The Stateless MCP Spec Audited My Compiler, and My Confirmation Gate Failed - What the 2026-07-28 MCP spec breaks in generated servers
- Your OpenAPI Spec Is Not an Agent Interface - Compiling governed MCP tool surfaces from OpenAPI and WSDL
- Beating FP16 with 4-bit Weights: A Portable W4A16 GEMM in Triton - 4-bit weight-only GEMM in pure Triton that runs on NVIDIA and AMD
- The Activation-Cone Blind Spot - Why prompt-time activation defenses miss prefilling attacks
→ More at subhadipmitra.com/blog
Data Platforms │ AI/ML Infrastructure │ Research
───────────────────────┼──────────────────────────┼─────────────────────
Context-Aware Pipelines│ LLM Evaluation │ Interpretability
ETLC / Context Stores │ Model Serving │ Activation Steering
Agent-Ready Platforms │ MLOps / LLMOps │ Formal Verification
Dynamic Context Engines│ Agentic Systems │ AI Safety
Singapore • Google Cloud • Schedule a call