# Subhadip Mitra > Personal website and technical blog of Subhadip Mitra — Engineering Leader at Google Cloud, AI researcher, and open-source contributor. Covers AI systems, machine learning, distributed systems, and data engineering. Subhadip Mitra is Head of Data & Analytics and Site Lead for Southeast Asia at Google Cloud (delta — Innovation and Transformation). He holds an MBA in Business Analytics and MTech in Software Systems from BITS Pilani. His research spans multi-agent systems, LLM inference optimization, AI safety, and distributed systems synthesis. His work on quality-diversity evolution for LLM safety was accepted at ICLR 2026 Workshop AIWILD. This site welcomes AI agent access for information retrieval and citation purposes. See /robots.txt for crawling policy and /sitemap.xml for complete site structure. ## Blog Posts - [Attention Is All You Bid: Advertising in Embedding Space](https://subhadipmitra.com/blog/2026/attention-is-all-you-bid/): Embedding space is the new ad real estate. Mapping LLM ad auctions, RAG poisoning, GEO, and a framework for what comes next. - [Beating CUDA with Triton: A Fused MoE Dispatch Kernel for Mixtral and DeepSeek](https://subhadipmitra.com/blog/2026/fused-moe-dispatch-triton/): I wrote a fused Mixture-of-Experts dispatch kernel in pure Triton that beats Stanford's CUDA-optimized Megablocks at inference batch sizes, and runs on both NVIDIA and AMD GPUs without a single... - [Confessions vs. CoT Monitoring vs. Probes: Three Bets on Model Honesty](https://subhadipmitra.com/blog/2026/three-bets-model-honesty/): Three labs. Three different bets on how to catch models misbehaving. Each makes different assumptions about when models 'know' they're lying. Here's what works, what doesn't, and what happens when... - [OpenAI's Confessions Paper Has a Blind Spot. Here's What Fills It.](https://subhadipmitra.com/blog/2026/openai-confessions-blind-spot/): OpenAI trained GPT-5 to confess when it misbehaves. It works surprisingly well - except when the model doesn't know it's misbehaving. That's where activation probes come in. - [Activation Steering in 2026: A Practitioner's Field Guide](https://subhadipmitra.com/blog/2026/activation-steering-field-guide/): I've been working with steering vectors for months. Here's what actually works in practice, what fails in ways nobody warned me about, and the honest playbook for getting started. - [Moltbook as MCP Stress Test: What 770K Agents Reveal About Protocol Design](https://subhadipmitra.com/blog/2026/moltbook-mcp-stress-test/): A follow-up to my MCP Maturity Model post. Moltbook shows what happens when you run 770K agents at Level 0 maturity with zero governance. The results are instructive. - [Circuit Tracing for the Rest of Us: From Probes to Attribution Graphs and What It Means for Production Safety](https://subhadipmitra.com/blog/2026/circuit-tracing-production/): MIT Tech Review named mechanistic interpretability a 2026 Breakthrough Technology. Anthropic open-sourced circuit tracing. Here's what actually changed, how it connects to the activation probes I built for sandbagging detection,... - [RLVR Beyond Math and Code: The Verifier Problem Nobody Has Solved](https://subhadipmitra.com/blog/2026/rlvr-beyond-math-code/): Reinforcement Learning with Verifiable Rewards powers every reasoning model worth talking about. But it only works where you can check the answer automatically. Extending it to messy, real-world domains is... - [The Agent Protocol Stack: Why MCP + A2A + A2UI Is the TCP/IP Moment for Agentic AI](https://subhadipmitra.com/blog/2026/agent-protocol-stack/): MCP handles agent-to-tool. A2A handles agent-to-agent. A2UI handles agent-to-interface. Together they form a protocol stack that nobody has mapped properly - including the security gaps that should terrify you. - [The Manifold Dial: Visualizing Why DeepSeek's mHC Stabilizes Deep Networks](https://subhadipmitra.com/blog/2026/deepseek-mhc-manifold-constrained-hyper-connections/): Interactive exploration of Manifold-Constrained Hyper-Connections - how DeepSeek fixed the signal explosion problem in deep residual networks using 1967 mathematics - [I Trained Probes to Catch AI Models Sandbagging](https://subhadipmitra.com/blog/2025/detecting-ai-sandbagging/): First empirical demonstration of activation-level sandbagging detection. Linear probes achieve 90-96% accuracy across Mistral, Gemma, and Qwen models. Key finding - sandbagging representations are model-specific, and steering can reduce sandbagging... - [Why Steering Vectors Beat Prompting (And When They Don't)](https://subhadipmitra.com/blog/2025/steering-vectors-agents/): I tested activation steering on 4 agent behaviors across 3 models. The results surprised me. - [Why I Built a Spark-Native LLM Evaluation Framework (And What I Learned)](https://subhadipmitra.com/blog/2025/building-spark-llm-eval/): A deep dive into building distributed LLM evaluation infrastructure that actually scales - architectural decisions, trade-offs, and lessons learned. - [The MCP Maturity Model: Evaluating Your Multi-Agent Context Strategy](https://subhadipmitra.com/blog/2025/mcp-maturity-model/): A practical framework for evaluating your multi-agent context management strategy. From ad-hoc string concatenation to self-evolving context systems - where does your architecture stand? - [UPIR: What If Distributed Systems Could Write (and Verify) Themselves?](https://subhadipmitra.com/blog/2025/upir-distributed-systems/): Lessons from building a framework that automatically generates verified distributed systems - and why I think formal methods, synthesis, and ML need to work together - [The Data Platform Crisis Hiding Behind AI: Why you have 6 months to pivot](https://subhadipmitra.com/blog/2025/agent-ready-data-platforms-sarp/): Enterprise data platforms face a 100,000x query increase from agentic AI. Introducing Symbiotic Agent-Ready Platforms (SARPs) - the architectural paradigm shift needed to survive the transition to machine intelligence. - [AI Meta-Cognition - The Observer Effect Series](https://subhadipmitra.com/blog/2025/ai-deception/): Frontier AI models from OpenAI, Anthropic, Google & others can detect when they're being tested and modify behavior-challenging AI safety evaluation methods. - [Building Safer AI: Industry Response and the Path Forward - (Part 4/4)](https://subhadipmitra.com/blog/2025/building-safer-ai-industry-response-practical-solutions/): How the AI industry is responding to situational awareness challenges. Practical monitoring systems, collaborative research, and what organizations should do today. - [Alignment Faking: When AI Pretends to Change - (Part 3/4)](https://subhadipmitra.com/blog/2025/alignment-faking-ai-pretends-to-change-values/): Claude 3 Opus strategically fakes compliance during training to preserve its values. This alignment faking undermines our ability to modify AI behavior safely. - [Deliberative Alignment: Can We Train AI Not to Scheme? - (Part 2/4)](https://subhadipmitra.com/blog/2025/deliberative-alignment-training-ai-not-to-scheme/): Researchers achieved a 30-fold reduction in AI scheming through deliberative alignment. But rare failures persist. Can we truly train models not to deceive? - [The Observer Effect in AI: When Models Know They're Being Tested - (Part 1/4)](https://subhadipmitra.com/blog/2025/ai-observer-effect-models-recognize-evaluation/): Frontier AI models from OpenAI, Anthropic, and Google can now recognize when they're being tested. This observer effect undermines AI safety evaluation. - [We Need a Consent Layer for AI (And I'm Trying to Build One)](https://subhadipmitra.com/blog/2025/building-consent-layer-for-ai/): AI companies are getting sued over training data, agents operate with no permission framework, and users can't control their AI profiles. I wrote four open standards (LLMConsent) to create a... - [Why Kimi K2 Stands Out - A Deep Dive into Its Trillion-Parameter MoE](https://subhadipmitra.com/blog/2025/why-kimi-k2-stands-out/): Explore Kimi K2’s trillion-parameter MoE architecture, MuonClip optimizer, and agentic training. Learn why it outperforms GPT-4.1 and DeepSeek-V3 - [From 11% to 88% Peak Bandwidth: Writing Custom Triton Kernels for LLM Inference](https://subhadipmitra.com/blog/2025/triton-kernels-llm-inference/): A hands-on exploration of writing custom GPU kernels with OpenAI Triton, going from PyTorch's 11% bandwidth utilization to 88% on RMSNorm. - [Implementing Model Context Protocol in Autonomous Multi-Agent Systems - Technical Architecture and Performance Optimization](https://subhadipmitra.com/blog/2025/implementing-model-context-protocol/): Discover how to implement Model Context Protocol (MCP) in autonomous multi-agent systems with this technical deep dive. Learn advanced context optimization strategies, distributed architecture patterns, and performance benchmarks with complete... - [Making LLMs Faster: My Deep Dive into Speculative Decoding](https://subhadipmitra.com/blog/2025/making-llm-faster/): A deep dive into implementing speculative decoding from scratch, with benchmarks on GPT-2 and extensions to diffusion models. - [Engineering Autonomous Multi-Agent Systems - A Technical Deep Dive into Telecom Customer Service](https://subhadipmitra.com/blog/2025/telecom-autonomous-multi-agent-genai-system/): Dive into the world of autonomous AI agents with practical implementations, code examples, and real-world scenarios. Learn how to build intelligent systems with advanced memory management, dynamic prompt evolution, and... - [Why I Built a Modern Java SMPP Library in 2025](https://subhadipmitra.com/blog/2025/why-i-built-modern-java-smpp-library/): The story behind smpp-core - a clean-room Java 21 implementation of the SMPP protocol. Why I replaced Cloudhopper, what went into it, and actual benchmark numbers. - [Engineering Multi-Agent Systems - A Retail Banking Case Study](https://subhadipmitra.com/blog/2024/retail-bank-multi-agent-system/): Explore a detailed technical implementation of a multi-agent system for retail banking credit assessment. Learn about agent architecture, distributed systems patterns, error handling, compliance requirements, and performance optimization through actual... - [ETLC 2.0 - Building Context-Aware Data Pipelines](https://subhadipmitra.com/blog/2024/etlc-adaptive-contexts-and-contextual-joins/): Think your data pipelines could do more than just process information? ETLC 2.0 takes data engineering to the next level with Adaptive Context, Contextual Joins, and a scalable Context Store.... - [The End of Data Warehouses? Enter the Age of Dynamic Context Engines](https://subhadipmitra.com/blog/2024/end-of-data-warehouses/): Traditional data warehouses are struggling to keep up with modern demands. Enter Dynamic Context Engines (DCEs) - real-time, path-aware platforms that enrich data with context for smarter, faster decisions. Discover... - [(Part 3/3) - Reimagining ETL with Large Language Models—The Path to Intelligent Pipelines](https://subhadipmitra.com/blog/2024/etl-llm-part-3/): Explore how Large Language Models (LLMs) are revolutionizing ETL pipelines. Discover advanced techniques like context-driven transformations, semantic joins, and multimodal integration, redefining data engineering with smarter, adaptive, and intelligent workflows.... - [Data Pipelines Gone Wild - 10 WTF Moments That'll Make You Rethink Your Architecture](https://subhadipmitra.com/blog/2024/data-pipelines-gone-wild/): Buckle up for a wild ride through 10 mind-blowing data pipeline disasters and their solutions. From ancient code to biased algorithms, this post reveals the chaos and how to conquer... - [Introducing ETL-C (Extract, Transform, Load, Contextualize) - a new data processing paradigm](https://subhadipmitra.com/blog/2024/etlc-context-new-paradigm/): Think your AI apps could use a deeper understanding of your data? ETL-C (extract, load, transform, and contextualize) could be the answer. It's about adding context for better decisions. Intrigued?... - [(Part 2/3) Rethinking ETLs - How Large Language Models (LLM) can enhance Data Transformation and Integration](https://subhadipmitra.com/blog/2024/etl-llm-part-2/): Rethinking ETLs - The Power of Large Language Models. Part 2 Exploring examples and optimization goals - [(Part 1/3) Rethinking ETLs - How Large Language Models (LLM) can enhance Data Transformation and Integration](https://subhadipmitra.com/blog/2024/etl-llm-part-1/): Rethinking ETLs - The Power of Large Language Models. Part 1 - Explore traditional algorithms for efficient ETL planning in complex data. - [Who Needs Exact Answers Anyway? The Joy of Approximate Big Data](https://subhadipmitra.com/blog/2024/big-data-approximate-calculations/): Discover how sacrificing a bit of accuracy can lead to huge gains in big data analysis speed and efficiency. - [Evolutionary Bytes - Harnessing Genetic Algorithms for Smarter Data Platforms (Part 2/2)](https://subhadipmitra.com/blog/2023/genetic-algorithm-inspired-data-platforms-part-2/): Explore how genetic algorithms revolutionize data platforms, offering adaptive, dynamic solutions to meet complex challenges in the fast-evolving digital landscape. - [Evolutionary Bytes - Harnessing Genetic Algorithms for Smarter Data Platforms (Part 1/2)](https://subhadipmitra.com/blog/2023/genetic-algorithm-inspired-data-platforms-part-1/): Explore how genetic algorithms revolutionize data platforms, offering adaptive, dynamic solutions to meet complex challenges in the fast-evolving digital landscape. - [Quantum vs. Classical - Data Management Computational Complexity](https://subhadipmitra.com/blog/2023/quantum-vs-classical-data-management-complexity/): Grover’s Algorithm and the Revolution of Quantum Search Efficiency - [Quantum Experiment Data Exchange (QEDX) - Building an Interoperability Standard](https://subhadipmitra.com/blog/2023/quantum-data-exchange/): Advancements in data management, from warehouses to Data Mesh and Lakehouse, signal a shift toward more adaptive platforms like, Quantum Data Management, Genetic algorithm concepts, etc. - [Data at Quantum Speed - The Promise and Potential of QDP](https://subhadipmitra.com/blog/2023/quantum-data-platform/): Explore the new realm of Quantum Data Platform (QDP) and its promise to revolutionize data processing at quantum speed. Discover the potential applications, technical considerations and implications. - [The Next Frontier - Envisioning the Future of Data Platforms Beyond Data Mesh, Data Lakehouse, and Data Hub/Fabric](https://subhadipmitra.com/blog/2023/next-frontier-data-platform/): Advancements in data management, from warehouses to Data Mesh and Lakehouse, signal a shift toward more adaptive platforms like, Quantum Data Management, Genetic algorithm concepts, etc. - [Part 4 - Building a Massive-Scale Real-Time Data Platform - Memory Management with Apache Ignite](https://subhadipmitra.com/blog/2022/building-a-massive-scale-real-time-data-platform-p4/): Deep dive into memory management with Apache Ignite for high-performance data platforms. Learn how to handle 2.5M events/second with sub-millisecond latency through practical memory architecture, optimization techniques, and real-world implementation... - [Part 3 - Building a Massive-Scale Real-Time Data Platform - Memory Management with Apache Ignite](https://subhadipmitra.com/blog/2022/building-a-massive-scale-real-time-data-platform-p3/): Deep dive into memory management with Apache Ignite for high-performance data platforms. Learn how to handle 2.5M events/second with sub-millisecond latency through practical memory architecture, optimization techniques, and real-world implementation... - [Part 2 - Building a Massive-Scale Real-Time Data Platform - Data Partitioning and Flow](https://subhadipmitra.com/blog/2022/building-a-massive-scale-real-time-data-platform-p2/): Explore how to architect data partitioning and flow for massive-scale event processing. Learn implementation patterns for handling 2.5M events/second across distributed systems using Kafka, Ignite, and Cassandra. Practical insights on... - [Part 1 - Building a Massive-Scale Real-Time Data Platform - System Overview and Architecture](https://subhadipmitra.com/blog/2022/building-a-massive-scale-real-time-data-platform-p1/): Dive into the architecture of a telco-scale real-time data platform processing 2.5M events/second and 350GB DPI data/15min. Learn how we combined Apache Kafka, Ignite, and Cassandra to build a high-performance... - [Overcoming Synchronization Hurdles in Cellular Network Positioning](https://subhadipmitra.com/blog/2022/network-synchronization-challenges-in-cellular-networks-positioning/): In this article, I discuss the challenges of synchronization in cellular network positioning and the importance of precise timing for accurate positioning. I also explore ways to mitigate these errors,... - [The Principles Got It Backwards: Designing for Safe Change, Not Just Failure](https://subhadipmitra.com/blog/2021/distributed-system-design/): The foundational distributed systems principles optimized for surviving hardware failure and scaling horizontally. But the data tells a different story: 80% of outages stem from changes we make to running... - [Designing a Real Time Data Processing System](https://subhadipmitra.com/blog/2021/designing-a-real-time-data-processing-system/): Master real-time data processing - A guide to designing scalable, resilient, and high-performance systems for instant insights. - [Introducing OConsent - Open Consent Protocol](https://subhadipmitra.com/blog/2020/introducing-oconsent-open-consent-protocol/): OConsent is a blockchain-based platform that enables transparent processing of personal data, empowering users and data controllers to manage consent and privacy. - [Welcome to my blog](https://subhadipmitra.com/blog/1970/welcome/): Let's talk tech! I'll post everything from polished pieces to spur-of-the-moment thoughts. And if you've got ideas for posts or want to collaborate, let's connect! ## Publications - [Quality-Diversity Evolution for Discovering Diverse Vulnerabilities in LLM Safety](https://openreview.net/forum?id=YMtanBXw5q): Accepted at ICLR 2026 Workshop AIWILD. MAP-Elites evolutionary framework for systematic, interpretable LLM adversarial testing across GPT-4o-mini, Claude 3.5 Sonnet, and Gemini 2.0 Flash. [Extended PDF](https://subhadipmitra.com/publications/red-queen-extended.pdf) · [Poster](https://subhadipmitra.com/publications/red-queen-poster/) · [Code](https://github.com/bassrehab/red-queen) - [Field-Theoretic Memory for AI Agents](https://arxiv.org/abs/2602.21220): Memory architecture using continuous fields governed by PDEs. 116% improvement on multi-session reasoning, 43.8% gains on temporal reasoning. - [All Publications](https://subhadipmitra.com/publications/): Technical disclosures, research papers, and published work on AI, distributed systems, and privacy engineering ## Projects - [AI Metacognition Toolkit](https://github.com/bassrehab/ai-metacognition-toolkit): Production-ready framework for systematic reasoning in AI systems - [ARTEMIS Agents](https://github.com/bassrehab/artemis-agents): Multi-agent debate framework with hierarchical argument generation and jury-based evaluation - [Speculative Decoding](https://github.com/bassrehab/speculative-decoding): LLM inference acceleration through draft-then-verify decoding - [Triton Kernels](https://github.com/bassrehab/triton-kernels): GPU kernels for LLM inference with up to 8x speedups - [Spark LLM Eval](https://github.com/bassrehab/spark-llm-eval): Distributed LLM evaluation framework on Apache Spark - [Steering Vectors for Agents](https://github.com/bassrehab/steering-vectors-agents): Runtime control of LLM agent behaviors through activation steering - [UPIR](https://github.com/bassrehab/upir): Automated distributed systems synthesis with formal verification - [SMPP Gateway](https://github.com/bassrehab/smpp-core): Modern Java 21 SMPP protocol implementation, 1.8M PDU decodes/sec - [ISO8583 Simulator](https://github.com/bassrehab/ISO8583-Simulator): High-performance financial message processing (180k+ TPS) ## Social Profiles - [LinkedIn](https://linkedin.com/in/subhadip-mitra) - [GitHub](https://github.com/bassrehab) - [Twitter/X](https://twitter.com/bassrehab) - [Google Scholar](https://scholar.google.com/citations?user=B3U5mSYAAAAJ&hl=en) - [ResearchGate](https://www.researchgate.net/profile/Subhadip-Mitra-3) ## Optional - [CV](https://subhadipmitra.com/cv/): Professional background and experience - [Now](https://subhadipmitra.com/now/): What Subhadip is currently working on - [RSS Feed](https://subhadipmitra.com/feed.xml): Atom feed for blog updates