I’m Subhadip Mitra. I embed with enterprises and ship production AI agents into their live systems, then publish the research and practitioner’s notes behind them: multi-agent systems, inference optimization, and AI safety.
MAP-Elites evolves interpretable attacks across 36 strategy-encoding combinations. No attack on Claude 3.5 Sonnet scored above ambiguous; GPT-4o-mini and Gemini 2.0 Flash each had attacks at 0.8.
Prompt-time activation defenses are structurally blind to prefilling attacks. Probing the first generated tokens catches them.
At inference batch sizes on NVIDIA A100 and AMD MI300X, with 35% less global memory traffic and no CUDA.
Linear probes on hidden states catch deliberate underperformance across Mistral, Gemma and Qwen, before any output is generated.
DeepSeek open-sourced DeepInfer and disclosed 155 trillion tokens a day at $38 per GPU-hour. The arithmetic says the cache it kept is the business.
Read the essayFrom API specifications to governed MCP tool surfaces. With provenance, policy, and human approval built in.
Portable GPU kernels for inference. Exploring the practical limits of memory bandwidth and expert routing.
A framework for structured debate between AI agents, with jury evaluation and integrated safety monitoring.
Peer-reviewed papers, technical disclosures, and preprints spanning AI safety, multi-agent systems, inference optimization, distributed systems, and privacy engineering.
View publicationsSome arguments are easier to poke than to read. These are interactive versions of claims I have made in the writing.
Open the instruments
Before Delta, I spent five years in Google Cloud’s professional services across Southeast Asia, rising from consultant to Head of Data & Analytics and Site Lead. I built the Data & Analytics practice there from zero across six countries. I also started an AI agent program for a 700+ person organization, and wrote the SDK it runs on.