Subhadip Mitra · Google Cloud

Data platforms. AI systems. The infrastructure between them.

delta at Google Cloud, leading Data & AI innovation and transformation across Southeast Asia. I publish research on multi-agent systems, inference optimization, and AI safety, and write the practitioner's notes behind it, from inside real production systems.

21 publications · 9 packages LatestICLR 2026 Spark-LLM-Eval

01 Now

Building

ICLR 2026 paper on LLM safety, and compute primitives for orbital environments.

Exploring

New ventures at the frontier; conversations with research labs and founders.

Writing

Fused MoE kernels, circuit tracing in production, and bets on model honesty.

What I'm up to →

02 Latest

§ AI-SAFETY · ESSAY · Jul 2026

The Activation-Cone Blind Spot, or Why Your Jailbreak Defense Can't See Prefilling

Prompt-time activation defenses stop GCG and AutoDAN cold, then fail half the time against prefilling. New paper: why that blind spot is structural, and how a linear probe on the first generated tokens closes it - AUR...

Read the essay →

03 Selected writing

INTERPRETABILITYJul 2026 What Runtime Interpretability Actually Costs, Part 1: The Case for Measuring ItEveryone assumes activation probes are too expensive to run in production. I ran the numbers ... DEEPSEEKJun 2026 DeepSeek DSpark: Speculation Is a Scheduling ProblemWhat DeepSeek's DSpark and DeepSpec release actually changes for LLM inference: suffix decay,... AGENTSJun 2026 Loop Engineering: Convergence Is Not CorrectnessA loop always converges. That it reached a stable 'done' state tells you nothing about whethe... LLMApr 2026 Attention Is All You Bid: Advertising in Embedding SpaceEmbedding space is the new ad real estate. Mapping LLM ad auctions, RAG poisoning, GEO, and a... DEEP-LEARNINGMar 2026 Beating CUDA with Triton: A Fused MoE Dispatch Kernel for Mixtral and DeepSeekI wrote a fused Mixture-of-Experts dispatch kernel in pure Triton that beats Stanford's CUDA-...

All essays in the archive →

04 Research focus

/ 01

Multi-agent systems

Coordination, debate, and verification for AI that runs unattended.

/ 02

Inference optimization

Fused MoE dispatch, speculative decoding, custom Triton kernels.

/ 03

AI safety & interpretability

Circuit tracing, activation steering, sandbagging detection.

/ 04

Distributed systems

Formal synthesis and correct-by-construction architectures.

05 Instruments

INFERENCEInteractive Memory-bandwidth rooflineMemory-bound or compute-bound, and how KV cache and kernel efficiency move the wall. ADVERSARIALInteractive Quality-diversity archiveWhy MAP-Elites maps diverse vulnerabilities that a single objective misses.

All instruments →

06 Selected publications

2026

Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense

arXiv

PDF →

2026

Cross-Generational Transfer of Adversarial Attacks Reveals Non-Monotonic Safety Alignment in LLMs

arXiv

PDF →

2026

Cross-Platform Fused MoE Dispatch in Triton: Portable Expert Routing Without CUDA

arXiv

PDF →

2026

Spark-LLM-Eval: A Distributed Framework for Statistically Rigorous Large Language Model Evaluation

arXiv

PDF →

2026

Quality-Diversity Evolution for Discovering Diverse Vulnerabilities in LLM Safety

ICLR 2026 Workshop AIWILD

PDF →

All publications →