About Rajkiran Panuganti
CEO, Cerebro Labs Inc. · Agentic Harness · Web-based agents · Former Head of AI, Krutrim · Microsoft Copilot · Founding team of Bing
CEO of Cerebro Labs Inc., building Agentic Harness to make web-based agents as powerful as terminal-based agents.
LLMs from scratch
Led the Krutrim team that trained India's first LLM, including a 7B model pretrained from scratch and a 12B follow-on model.
Agentic Harness
Building Agentic Harness at Cerebro Labs Inc. to make web-based agents as powerful and reliable as terminal-based agents.
Copilot and Bing
Spent 14 years at Microsoft across Microsoft Copilot, the founding team of Bing, Cortana, Azure ML, privacy, evaluation, and Responsible AI.
Rajkiran Panuganti is CEO of Cerebro Labs Inc., where he is building Agentic Harness to make web-based agents as powerful and reliable as terminal-based agents. He is an AI technology leader with two decades of experience building large-scale AI systems, and is among the few AI leaders in India with hands-on experience training LLMs from scratch.
As Head of AI at Krutrim, Rajkiran led AI research and product efforts across foundation models, agentic assistants, multimodal systems, Indian-language AI, guardrails, document intelligence, ASR/TTS, coding copilots, personalization, and enterprise GenAI applications. His team trained India's first LLM at Krutrim, including the 7B model pretrained from scratch, and later launched a 12B model.
At Cerebro Labs Inc., he is building Agentic Harness for web-based agents: systems that can reason over enterprise context, operate websites, orchestrate tools, follow operational constraints, and execute multi-step work with the same discipline users expect from terminal-based agents.
Before Krutrim, Rajkiran spent 14 years at Microsoft, where he worked on Microsoft Copilot and Microsoft 365 Copilot Chat, the AI assistant that helps enterprise users retrieve information across their files, emails, and chats. His earlier work at Microsoft included Bing, where he was part of the founding team, Cortana's natural language understanding stack, and the experimentation platform that became Azure Machine Learning.
His Bing work included Sort Rank, the algorithm that determined the relative importance of every webpage in the Bing index, and the Names Ranker that powered web search ranking for queries containing person entities. He holds patents from his time at Microsoft, including work on search refiner generation. Before Microsoft, he was a software engineer at Google in the platform infrastructure team.
Today, Rajkiran is building Agentic Harness at Cerebro Labs Inc. Alongside that, he advises AI startups including Codemod, Nelumbium Capital, and AppAxon, helping them navigate the practical and architectural challenges of building production Generative AI systems. He also conducts independent research on mechanistic interpretability. His paper CircuitProbe (arXiv, 2026) introduces a method for detecting reasoning circuits inside transformer models using activation statistics, replacing brute-force layer search with a process roughly 1000x faster.
Alongside his research and advisory work, Rajkiran writes a weekly deep-dive newsletter analyzing what's actually happening at the frontier of AI: model releases, agent infrastructure, research trends, and the gap between what gets demoed and what actually ships in production. The same column runs on LinkedIn as My Thoughts on LLMs.
Rajkiran holds a PhD in Computer Science from The Ohio State University, where his research focused on enabling high-performance computing inside high-productivity languages like MATLAB. He earned his bachelor's in Electronics and Communications Engineering from the Indian Institute of Technology Bombay, where his thesis on computer vision was advised by Dr. Subhasis Chaudhuri.
He writes for engineers, researchers, and builders who want substance over hype.
Currently
- CEO · Cerebro Labs Inc.May 2026 — Present
Building Agentic Harness to make web-based agents as powerful and reliable as terminal-based agents.
- Advisor · CodemodJan 2024 — Present
Advising on AI strategy. Codemod builds developer tooling that automates undifferentiated work across massive codebases, including those produced by AI.
- Advisor · AppAxonJan 2024 — Present
Advising on Generative AI applications in cybersecurity.
- Strategic Advisor · Nelumbium CapitalMar 2023 — Present
Advising on the AI engine that powers their systemic-risk modeling. Nelumbium applies complexity economics to financial systems, with the premise that shocks to the system are not exogenous.
Previously
- Founding Head of AI · BlazelSep 2025 — May 2026
Led AI for agentic marketing at Blazel, a LinkedIn content and GTM platform that turns company-wide executive and team voices into posts, audience intelligence, lead enrichment, and CRM pipeline.
- Head of AI · KrutrimFeb 2024 — Feb 2026
Led AI research and product efforts across foundation models, agentic assistants, multimodal AI, Indian-language AI, document intelligence, guardrails, speech, coding copilots, personalization, and enterprise GenAI applications.
- Led the team that trained India's first LLM at Krutrim, including a 7B model pretrained from scratch and a 12B follow-on model.
- Built model and agentic AI systems that connected foundation-model capability to real product workflows.
- Shipped Indian-language model capability including a world-class bi-encoder for vector databases and a novel tokenizer for phonetic Indian languages.
- Built and shipped document intelligence, guardrails, hallucination-resistant chatbot applications, ASR/TTS, LLaVA-style Vision LLMs, agent-builder infrastructure, and a VS Code coding copilot.
- Fine-tuned LLaMA 3.1 70B and LLaMA 3.2 90B for production use cases.
- Patents pending for phonetic tokenization and LLM-based user profile generation for shopping recommendations.
- Applied Science · Microsoft Copilot and Bing · Microsoft2009 — 2023
14 years at Microsoft across Microsoft Copilot, Microsoft 365 Copilot Chat, Bing, Cortana, Microsoft 365 Search, and Azure ML. Worked on LLM behavior, retrieval-augmented generation, evaluation metrics, privacy, and Responsible AI for the enterprise Copilot experience; earlier, was part of the founding team of Bing.
- Microsoft Copilot and Microsoft 365 Copilot Chat: deep work on LLM behavior, RAG, metric design, privacy and Responsible AI for the enterprise Copilot.
- Founding team of Bing: built core search-ranking and web-scale relevance infrastructure in the early Bing era.
- Cortana: NLP, intent and entity extraction, and the orchestration engine that later evolved into the orchestration layer powering BizChat.
- Bing Sort Rank: variant of PageRank that determined the relative importance of pages in the Bing index. Saved hundreds of millions in serving costs.
- Bing Names Ranker: built a person knowledge graph from the open web to power people-related search queries and disambiguation.
- Azure ML: built the experimentation workflow system from scratch, which matured into part of Azure Machine Learning.
- Two patents granted, including search refiner generation for applications.
- Software Engineer · Google2007 — 2008
Platform infrastructure. Performance tuning of widely used internal tools through scalable architecture redesigns.
- Research Assistant · Ohio Supercomputer CenterJul 2004 — Dec 2007
Research on high-productivity languages for parallel computing.
- Research Scientist (Internship) · Hewlett-PackardJun 2006 — Sep 2006
Developed High-Performance MATLAB for parallel computing.
- Engineering Intern · Sun MicrosystemsMay 2002 — Jun 2002
Built a bug tracking system.
Education
- PhD, Computer Science · The Ohio State University2003 — 2008
Specialized in enabling high-performance frameworks for machine learning workloads, including distributed systems, parallel computing, and NVIDIA CUDA. Thesis: enabling high-performance computing in high-productivity languages like MATLAB.
- Bachelor of Technology, Electronics and Communications Engineering · Indian Institute of Technology Bombay1999 — 2003
Thesis in computer vision, advised by Dr. Subhasis Chaudhuri. First introduction to machine learning.
Research
FABRIC: AI Financial Advisors Hallucinate More Than They Forget on Indian Markets
SSRN · 2026A benchmark of 204 verified Indian financial questions across six languages. Evaluates seven models on 16,000+ responses. Finds hallucination is the dominant failure mode, Indian-origin models do not perform better on Indian financial questions, and Hinglish outperforms pure Hindi. WebRAG dramatically reduces both hallucination and outdated errors.
CircuitProbe: Predicting Reasoning Circuits in Transformers via Stability Zone Detection
arXiv · 2026A method for predicting where reasoning circuits live inside transformer models by detecting stability zones in activation statistics. Replaces brute-force layer search with a process roughly 1000x faster.
MATLAB: A Language for Parallel Computing
International Journal of Parallel Programming · 2008Presents extensions to MATLAB for parallel computing including mexMPI for message passing and GAMMA for distributed shared memory, enabling high-performance computing while retaining MATLAB's productivity.
A Learning-Based Method for Image Super-Resolution from Zoomed Observations
IEEE Transactions on Systems, Man and Cybernetics, Part B · 2005A machine learning approach to image super-resolution using observations at different camera zoom levels.
Integrated Loop Optimizations for Data Locality Enhancement of Tensor Contraction Expressions
SC 2005 (Supercomputing) · 2005Compiler optimizations for improving data locality in tensor contraction computations used in quantum chemistry.
Cache Miss Characterization and Data Locality Optimization for Imperfectly Nested Loops on Shared Memory Multiprocessors
IPDPS 2005 · 2005Analysis and optimization of cache performance for irregular loop structures on shared-memory parallel systems.
Super-Resolution Imaging: Use of Zoom as a Cue
Image and Vision Computing · 2004Obtaining high-resolution images of a scene from observations at different camera zoom levels. B.Tech thesis work at IIT Bombay, advised by Dr. Subhasis Chaudhuri.
A High Productivity Framework for Parallel Data Intensive Computing in MATLAB
PhD Thesis, The Ohio State University · 2008Presents mexMPI, GAMMA, and LA: three frameworks enabling parallel computing directly in MATLAB for high-performance workloads while retaining MATLAB's productivity. Advisor: P. Sadayappan.