Harsh Maheshwari

📢 New Paper Alert! Dr. Aravind Reddy, Principal Project Scientist at CeRAI, WSAI, IIT Madras, recently co-authored a paper that was accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026, one of the leading computer vision conferences. This was a joint work with Dr. Konda Reddy Mopuri and Mr. Saumyaranjan Mohanty, Department of AI, IIT Hyderabad. The paper, titled “DiRe: Diversity-promoting Regularization for Dataset Condensation”, proposes a simple, widely applicable diversity regularizer that can be plugged into existing dataset condensation methods. They show that DiRe consistently improves both test accuracy and dataset diversity across multiple benchmarks (CIFAR-10/100, Tiny-ImageNet, ImageNet-1K) and condensation methods, using principled metrics such as Coverage, Vendi Score, and CLIP-based coverage. Conceptually, the work supports CeRAI’s goals in Responsible AI by: ✅ enabling more data- and energy-efficient training via better synthetic datasets, ✅ making dataset diversity a quantifiable property rather than a vague claim, and ✅ strengthening the use of privacy-friendly synthetic data that remains robust across multiple architectures. Paper link: https://lnkd.in/gk2q5tpW Link to the code for the paper: https://lnkd.in/g7BtMQvw Balaraman Ravindran | Aravind Reddy | Centre For Responsible AI (CeRAI) | Wadhwani School of Data Science and AI, IIT Madras | Konda Reddy Mopuri | Saumyaranjan Mohanty | Department of AI, IIT Hyderabad | Robert Bosch Centre for Data Science and Artificial Intelligence (RBCDSAI) | IBSE: Centre for Integrative Biology and Systems medicinE | Indian Institute of Technology, Madras

49

🔥 Real Madrid spent $80M on Ronaldo, and Meta spent $100M to land Jiahui Yu from OpenAI. But the real game-changer? Trapit Bansal, an IIT Kanpur grad who went from Accenture analyst to IISc researcher to UMass Amherst PhD. He’s worked at Google, Facebook, Microsoft, and OpenAI, helped build the o1 reasoning model, and now joins Meta’s Superintelligence Labs with a rumored $100M payday. Meta didn’t stop there. They’ve poached a wave of core OpenAI talent: - Shengjia Zhao (ChatGPT and GPT-4 co-creator) - Jiahui Yu (o3, o4 mini, GPT-4.1) = Shuchao Bi and Hongyu Ren (GPT-4o creators) Meta is building its own $100M dream team. This isn’t just a hiring spree. It’s a strategic knockout move to drain OpenAI of its key people and leap ahead in AGI. 🧠 What this means: Meta is rewriting the roster with superstar engineers OpenAI risks serious talent loss even as it focuses on retention The AI battlefield now includes billion-dollar sign-on bonuses Bottom line: From an Accenture cubicle to Meta’s Superintelligence Labs, this high-stakes transfer saga could reshape the future of AI. Who’s on your fantasy team?

22

2 Comments

Recently started CS772A: Probabilistic Machine Learning by Piyush Rai (IIT Kanpur), and the first lecture reinforces something every ML researcher eventually realizes: Point estimates are not enough. Link: https://lnkd.in/gvSZZYwA Modern ML problems demand principled uncertainty modeling, not just predictions. Core research takeaways: Learning in ML can be framed as estimating probability distributions, not just functions Predictive inference: p(y∗∣x∗,D) Generative modeling: p(x∗∣D) The sum rule (marginalization) and product rule form the backbone of: Bayesian inference Latent variable models Variational methods and MCMC *Clear decomposition of uncertainty: Epistemic uncertainty → model/parameter uncertainty (reducible with data) Aleatoric uncertainty → inherent data noise (often irreducible) *Why Bayesian methods naturally provide: Posterior over parameters p(θ∣D) Predictive distributions with calibrated uncertainty *Conceptual links to research areas like Active learning Bayesian optimization Robust decision-making under uncertainty The lecture also highlights the generative vs.. discriminative modeling trade-off and why generative models are foundational to modern advances in representation learning and simulation. For anyone aiming at ML research, PhD work, or industry research labs, probabilistic ML is not optional—it’s the language underlying reliable and interpretable AI systems. Looking forward to deeper dives into Bayesian inference, generative models, and uncertainty-aware learning. #MachineLearningResearch #ProbabilisticMachineLearning #BayesianInference #UncertaintyQuantification #GenerativeModels #CS772A #IITKanpur #MLTheory

11

Jamia, IIT-M, IITH, Amity among institutions using AI, ML models to decode gene mutations, identify targets, and improve survival predictions and treatments, writes Aranya Shankar. Read more: https://lnkd.in/d-qKursX #AI #Research #Cancer #IIT #Healthcare

1

At IIT Bombay this weekend, one team proved that powerful AI does not have to mean reckless AI. We are extremely proud to announce that team Claudius Maximum, Oditi Rathi, Kiara Thapar and Denver Zuzarte, won Intent Intelligence Weekend, hosted at IIT Bombay by ArmorIQ in collaboration with WNCC and the AIC community. Their project stood out for one reason: architecture with discipline. They built a policy-enforced travel booking system where AI can suggest flights and hotels but cannot execute anything without passing strict backend rules. What impressed us: • A 5-stage pipeline turning “book me a flight” into a validated transaction • Live pricing APIs with real-time budget and compliance checks • A two-agent setup separating reasoning from execution The philosophy was simple: Let AI think freely. Never let it act freely. Huge congratulations to the team for raising the bar at IIT Bombay. Special thanks to Lokit Kumar Paras for judging and supporting the event. The future of AI will be built on campuses. Secured by design. #IITBombay #ArmorIQ #IntentIntelligence #AgentSecurity #AI

76

4 Comments

We are excited to share an update from GAHNA, India's Sovereign AI initiative (Generative Architecture for Hyperlocalized Neural Assistants). Please follow this page for more updates on GAHNA initiative. Built under the vision of Dr. Utpal Chakraborty(PhD), GAHNA is creating transparent, auditable, and domain-specific Small Language Models (SLMs) that will empower industries and governments to have full control over AI. Two of our first models are almost ready for making an impact: 🔹 GAHNA UPI-ODR – a specialized model for seamless UPI online dispute resolution and chargeback management in the BFSI sector. 🔹 GAHNA CyberMind – an AI-driven cybersecurity assistant designed to provide real-time threat intelligence, mapping to MITRE ATT&CK & D3FEND, and delivering auditable insights for SOCs and national security. This is just the beginning. With GAHNA, we are committed to building a sovereign AI ecosystem that is transparent, reliable, and future-ready. #GAHNA #AI #SovereignAI #ResponsibleAI #ExorionAI #UPI #GAHNAUPI #Cybersecurity #GAHNACyberMind

18

Here is my 4th research paper (PARAM), which showcases LLMs' ability to provide prescriptive recommendations based on industry machinery user manuals (PDFs). This work continues our research on LAMP, for which the abstract has been accepted at an IISc Bangalore conference and the full paper is currently under review. In LAMP, we demonstrate whether LLMs can outperform traditional machine learning algorithms in numerical anomaly prediction. In LAMP (LLM-Aided Machine Prognosis), we analyzed ball bearing vibrational data to predict anomalous behavior. Based on those forecasts, PARAM (Prescriptive Agents based on RAG for Automated Maintenance) provides recommendations to resolve potential breakdowns. During this research, we compared multiple language models and evaluated their performance across different problem sets. Thanks to co-author Anupam Purwar for his consistent support throughout this work. The findings are promising, and we would welcome industry partners to test and apply this work in real-world applications, as it holds immense potential. PARAM paper arXiv - https://lnkd.in/g6aVsD8B Video presentation generated using SARAL platform - we thank the development team (Ponnurangam Kumaraguru (PK), Sai Ganesh, Vishnu Sathwik, Lakshmanan Nataraj, Rahul Sundar) for this innovative research dissemination tool. #DemocratiseResearch #ANRF #AI4ALL #Saral #Industryresearch

33

14 Comments

ayuRAG: My UCSC chatbot experiment Been deep in RAG research while building a University of California, Santa Cruz -specific chatbot, experimenting with Agentic RAG workflows like Adaptive RAG, Corrective RAG, Self-RAG, RAT-RAG, RARE-RAG, multi-agent systems,ReAct frameworks etc. Each has unique strengths that taught me something valuable. The result? Our hybrid, confidence-driven approach : ayuRAG (pun intended), hit 85.4% on RAGAS’s faithfulness + answer relevancy (vs. 81.2% for the next best), delivered strong BLEU and ROUGE-L scores, and ranked top 4 on the remaining RAGAS metrics(Context Precision and Context Recall). Remarkably, even our local-only variant still secured second place; outperforming methods that rely on web search to augment their knowledge. More importantly, I learned that the best workflow is usually a custom blend, not a copy-paste solution. ayuRAG workflow has 2 main components: 1. Retrieval tool with web search enabled - it can rewrite queries for efficient vector search retrieval with k tries (k = 3 in our case) and a boolean flag for web search. When enabled, it performs web search k times while rewriting the query (but we need the original query to write every new query or else information can be lost). This tool also has document reranking and document relevancy guardrails. 2. Agent - First calls the retrieval tool, then gets the answer. If it can answer with confidence, it returns the result. Else, we use Chain-of-Thought (CoT) to split the question into atomic steps and then call the retriever tool for each sub-thought before answering. Simple questions get direct answers without web search, complex ones get decomposed into manageable sub-problems. The workflow is simple yet powerful - the real magic happens in prompt engineering in these components! Key insights from building this: 1. Domain trumps architecture - Study your dataset and design your retriever based on the specific use case. Tune chunk size (e.g., 200–500 tokens), chunk overlap (20–50 tokens), and similarity scoring (e.g., cosine with dense embeddings). 2. Prompt engineering is undervalued - Better prompts often outperform complex agent orchestration by significant margins. 3. Reranking prevents hallucination - Too many documents in context overwhelms models. 4. Human queries NOT EQUAL Vector queries - "How do I apply?" won't match "university enrollment application procedures." Rewrite for embedding similarity and even for web search. 5. Cot doesn’t always mean it will make things better. If the query is simple, doing COT is an overkill. Would love to hear your thoughts! Which workflow patterns have worked best in your domain? #RAG #PromptEngineering #AgenticWorkflows #FAISS #LangChain #LangGraph #AdaptiveRAG #SelfRAG #VectorDatabases #DocumentRetrieval #LLMApplications #AIEngineering #MLEngineering #RAGSystems #ProductionAI #MachineLearning #OpenAI #ReAct #ChainOfThought #AIOptimization #NLP #Agents #AIAgents

54

4 Comments

#Delhi’s air is not uniformly polluted — and IIT Kanpur’s latest findings show why averages mislead policy. Using a mobile atmospheric laboratory, advanced chemical instrumentation, dense sensor networks and #AI, Indian Institute of Technology, Kanpur mapped #pollution at two contrasting locations in #Delhi. 🚨 The results are stark: • At #AnandVihar, PM2.5 averaged ~63 µg/m³ even in summer, driven largely by local road dust, traffic and combustion, with black carbon frequently exceeding 5 µg/m³. • Just 30 km away in #Dwarka, PM2.5 dropped to ~38 µg/m³, dominated by secondary aerosols transported regionally, with low black carbon — a background, not hotspot, environment. • In Dwarka, over 80% of PM2.5 formed in the atmosphere, not from direct emissions. The message is clear: #Delhi does not have one pollution problem — it has many, varying by place and time. IIT Kanpur’s AI-enabled source apportionment work, already achieving >90% accuracy in earlier deployments, points to a future of hyper-local, real-time, enforceable air-quality action. Read the full article here 👉 https://lnkd.in/gpJfgmqQ Sachchida Nand Tripathi Kotak School of Sustainability - IIT Kanpur #IITKanpur #KotakSchoolOfSustainability #CleanAir #AirPollution #DelhiAir #PM25 #AirQuality #EnvironmentalScience #AIforGood #DataDrivenPolicy #UrbanSustainability

10

Our paper "Impact of mode switching on Short Transfers over CoAP for D2D-enabled V2V Communication: A Study on Realtime Inference Using ANN Splitting" was accepted in the main track of 32nd National Conference on Communications (NCC) held in IIT Hyderabad this March, and was presented by my colleague Mr. Ashis Sau. The paper combines ANN Splitting with 5G PC5 interface and studies the impact on the application level throughput due to perceived loss caused by mode-switching between PC5 and Uu interfaces induced by mobility of nodes. Thanks to Amit Sethi for all the support.

35

🎉 Excited to Share Our Latest Research! 🎉 Our paper titled 💡 “FERMI-ML: A Flexible and Resource Efficient Memory-in-Situ SRAM Macro for TinyML Acceleration” authored by Mukul Lokhande, Akash Sankhe, S.V Jaya Chand and, and Dr. Santosh Kumar Vishvakarma, has been accepted for presentation at the International Conference on Microelectronics (ICM 2025) to be held in Cairo, Egypt, from December 14–17, 2025. This work presents FERMI-ML, a novel Memory-in-Situ SRAM macro aimed at enabling efficient and flexible TinyML acceleration directly within memory, improving both performance and resource efficiency. A special thanks to our supervisor, Dr. Santosh Kumar Vishvakarma for his continuous guidance, support, and inspiration throughout this work. 🙏 Looking forward to sharing our contributions and connecting with the global microelectronics research community at ICM 2025! #NSDCS_LAB_IIT_INDORE #TinyML #SRAM #ComputeInMemory #EdgeAI #ICM2025 #Research #VLSI #IITIndore #HardwareAcceleration #CIM

28

1 Comment

🌟 DMSE Research Gains Nationwide Media Attention 🌟 We are proud to share that pioneering research led by Prof. Nitya Nand Gosvami (DMSE, IIT Delhi) and Prof. N M Anoop Krishnan (Civil Engineering, IIT Delhi) has been published in Nature Communications. 📄 The study reports the development of an AI-driven autonomous agent capable of designing and conducting real scientific experiments in a laboratory environment, marking a significant advance in the integration of artificial intelligence with experimental materials and engineering research. 🔬 This interdisciplinary work bridges materials science, civil engineering, and AI, demonstrating how autonomous systems can accelerate scientific discovery by independently planning, executing, and learning from experiments. 🗞️ Featured across leading national media outlets: The Times of India The Hindu NDTV 🔗 Media coverage: https://lnkd.in/gurxUxES https://lnkd.in/gqjZV6F9 https://lnkd.in/gUTYYbcU 👏 Congratulations to the research team on this outstanding achievement, reinforcing DMSE and IIT Delhi’s leadership in high-impact, interdisciplinary research. 📘 Read the paper: 🔗 https://lnkd.in/gCm8gVki #DMSE #IITDelhi #NatureCommunications #AIinScience #AutonomousLabs #MaterialsScience #CivilEngineering #ResearchExcellence

51

Great news for the Neuromechanics Research Lab (NRL) IIT Delhi! Bhaskar Pandey, Ph.D. and Prof. Deepak Joshi have published their paper entitled "Investigating the Robustness of Fuzzy Deep Learning on Noisy Medical Images" in IEEE #Transactions on #Artificial #Intelligence. The study demonstrates that a #Deep #Neuro-#Fuzzy System maintains high classification performance across #multiple #biomedical #imaging #datasets, even under six diverse #noise conditions. Comparative evaluations show that it #outperforms #state-#of-#the-#art deep learning models on noisy data, highlighting its superior #robustness and #generalizability. The work also reveals the #susceptibility of both DNFS and conventional models to #adversarial perturbations. Find out more about their work at: https://lnkd.in/gxSVGdPa IEEE IEEE Computational Intelligence Society Neuromechanics Research Lab (NRL) IIT Delhi Centre for Biomedical Engineering, IIT Delhi Indian Institute of Technology, Delhi AIIMS (All India Institute of Medical Sciences, New Delhi)

37

2 Comments

Post 3 in my RL & GenAI series: How do you train an agent to make good decisions before it knows how things end? Something similar to how google maps reroute you mid-drive. That’s where Temporal Difference (TD) learning comes in. Unlike Monte Carlo methods, which wait until an episode finishes, TD learns step by step—updating predictions using its current guess of what happens next. This “bootstrapping” lets it learn faster, even if its estimates are still rough early on. There’s a meaningful tradeoff here: → TD: Fast, stable, but biased → MC: Slow, noisy, but unbiased In real-world systems, TD’s efficiency often wins out. And with tools like N-step TD, we can tune that balance between speed and accuracy—blending short-term estimates with long-term outcomes. In practice, TD’s efficiency makes it foundational. Algorithms like SARSA and Q-learning are built on this idea—and power everything from game-playing agents to recommendation systems. Find full writeup with math and equations here: https://lnkd.in/gwdkFQky (Check in comments for all my previous blog posts) #ReinforcementLearning #MachineLearning #Blogs #LearningDaily

18

1 Comment

Kolkata-based Assessli introduces the LBM, a self-supervised AI system that integrates genomics, neuropsychology, medical data, physiological signals, and real-time behaviour. Unlike conventional AI, LBM offers hyper-personalisation by understanding the biological and behavioural context of individuals. With 150,000+ active users and 20 million proprietary data points, the model adapts continuously, enabling applications across education, healthcare, and financial services. Teachers can tailor learning pathways, while privacy remains a priority through encrypted edge-computation and differential privacy. Backed by ISI, angel investors, and a $5M seed round, Assessli is expanding globally, targeting 1 million users by 2027. This innovation moves AI from answering questions to truly understanding the person asking them. Suraj Biswas | AVIK MODAK | Koyel Manna | Soudeep shaw | Om Pratap Singh | Chirantan Guha | RITWIKA DAS | Tiasha Majumder | Joy Biswas | Vineet Gupta | Saurav Gupta #Assessli #ArtificialIntelligence #LBM #HyperPersonalisedAI #AIInnovation #EdTech #HealthTech #FinTech #PrivacyByDesign #DataSecurity #AIForGood

82

2 Comments

"𝗜𝗡 𝗖𝗔𝗦𝗘 𝗬𝗢𝗨 𝗠𝗜𝗦𝗦𝗘𝗗 𝗜𝗧" In October 2025, the Data Analytics and Visualization (DAV) Team, IIT Bombay released a comprehensive report on research at IIT Bombay, showcasing the breadth and depth of research across the institute. What’s inside the report: • A detailed analysis of research performance and publication trends • A curated database of faculty research interests across multiple disciplines • A comparative study of research output and impact between IIT Bombay and other renowned institutes Explore the report to gain deeper insights into IIT Bombay’s research ecosystem and identify potential opportunities for collaboration and growth. #IITBombay #ResearchAtIITB #DataAnalytics #ResearchInsight #AcademicResearch #HigherEducation #DataVisualization

14

📌 Announcing the Inaugural #Workshop on Systems for Machine Learning Organized by the IISc ACM Student Chapter and IISc ACM-W Student Chapter. Are you fascinated by the tech that powers today’s smartest AI systems? Curious about what really happens behind the scenes of machine learning? We are delighted to announce the inaugural edition of this immersive 3-day event, taking place at the Indian Institute of Science (IISc), Bengaluru, from June 13–15, 2025. With the exponential growth of machine learning workloads, there’s a pressing need to develop robust systems, both hardware and software, that can scale, perform, and adapt. This workshop aims to highlight emerging trends, ongoing research, and practical approaches in designing systems that power modern ML applications. Whether you’re a systems enthusiast, an aspiring ML engineer, or just curious about what powers today’s AI, this is your chance to explore, engage, and get inspired! Please visit our website for speaker and registration details: https://lnkd.in/dZ5xpr74 📌 What’s in store: * Exclusive Autonomous Vehicles demo by NVIDIA experts. * Engaging keynotes and panel discussions featuring the country’s leading researchers from #IISc and #IITs, alongside industry veterans from top companies such as Microsoft, Hewlett Packard Enterprise and HARMAN International. * In-depth #tutorial sessions crafted on emerging topics. * Hands-on project with exciting prizes and interactive lab demos. * Connect and collaborate with like-minded peers, researchers, and professionals who are shaping the next generation of AI systems. 📚 Tutorial Sessions Lineup  1. Introduction & Overview: Systems in ML   → Explore the motivation, scope, and evolving landscape of systems for ML.  2. Hardware for ML – What’s Under the Hood?   → A look at GPUs, TPUs, ASICs, and the hardware stack driving ML performance.  3. Software & Profiler Basics for ML   → Understand system software, compilers, and profilers that optimize ML workflows.  4. Distributed ML – Getting Started   → Learn how ML models are trained at scale using distributed computing frameworks. 5. Case Study & Lab: Accelerating ML Workloads   → Get hands-on experience in optimizing ML pipelines for real-world scenarios. 6. Ensuring Reliability – Scalability & Fault Tolerance   → Discover design principles for building scalable and fault-tolerant ML systems. 7. Edge Computing – Deploying ML on the Go   → Learn to deploy ML models on edge devices like NVIDIA Jetson and Raspberry Pi. We have limited slots for only ~80 participants. So, hurry up and register. Registrations are open to #everyone who is excited about #machine learning, #systems architecture, and #AI infrastructure. Whether you're a beginner or a budding expert, this is your #opportunity to learn, build, and be inspired. #IISc #ACM #ACMW #MLSystems #MachineLearning #TechEvents #IIScBengaluru #TutorialSeries #MLWorkshop ACM India ACM, Association for Computing Machinery

85

1 Comment

LLM reasoning still has so many flaws that Stanford and Caltech researchers wrote an entire journal article categorizing them! 🤯 Fundamental failures – stemming from intrinsic architectural or training constraints – manifest across all reasoning types. For example, cognitive biases such as confirmation bias. Robustness issues cut across domains but are particularly well-studied in benchmark-based evaluations and social reasoning, where minor, semantically-preserving perturbations – such as reordering options in multiple-choice questions, renaming variables in code, or paraphrasing moral dilemmas – can lead to large and inconsistent shifts in model outputs. Application-specific limitations cluster in certain domains: for example, Theory of Mind instability in implicit social reasoning, and inability to generalize to novel Math Word Problem structures in formal reasoning. The paper is in Transactions on Machine learning Research, available on Arxiv: https://lnkd.in/gi4uw_BU One funny thing about the article is that there are so many co-authors on some LLM research papers that a single paper references may cover multiple pages and overall the references are nearly half the article length 😆 (Thank you to my two alma maters, from my undergrad and graduate degrees :) #LLM #LLMHype #AI #AIHype #AIBubble

8

2 Comments