Ramsri Goutham Golla

CraftifAI co-founded by Pratik Sharda and Yashwant Dagar, secured $3M in seed funding led by Ankur Capital to scale its GenAI platform. Targeting IoT, robotics, and automation, CraftifAI's silicon-agnostic platform accelerates embedded product development for edge devices. Their GenAI and Agentic AI workflow optimizes model deployment, supporting frameworks like GStreamer and ROS2. The funds will expand engineering and global reach. Pilot programs are underway with Indian OEMs in robotics, drones, and AI cameras, plus a US semiconductor firm, streamlining hardware development and reducing time to market. #GenAI #EdgeAI #Robotics #DeccanFounders #CraftiAI #AnkurCapital

41

🧾 Manifold Constrained Hyper Connections by DeepSeek Thanks to Uday Kamath, Ph.D. - who has a fantastic write-up on this. ❓What is it? A new technique that enables training higher quality LLMs with minimal increases in overhead. Modern Transformers have a highway “fast lane” to remember the initial “query”. To date, this has been limited to a small number of lanes. Previously, Deep Seek had shown that increasing the number of lanes improves quality…but that came with a lot of risk. Large-scale trainings could explode out of control, at immense capital cost. By constraining the manifolds, they shift the efficient frontier and increase reward with minimum increase in risk. A big breakthrough. 📈 Why it matters to business • Foundation models can have sizable jumps in quality with minimal costs changes. • The largest models can now be more ambitious without an increased fear of failed trainings. 🛠️ The technology details • Residual connection lanes are widened to incorporate more features. The new residuals have learned pre and post transformations. • In order to achieve stability, they incorporated the 1960s Sinkhorn-Knopp algorithm, which transforms the residual connection into a doubly stochastic matrix. • This markedly increased stability of training, measured by the Amax Gain Magnitude. • The team also devised some interesting innovations in memory management and gradient checkpointing in order to achieve these results. Uday’s Writeup https://lnkd.in/eSD8i9wP Paper https://lnkd.in/eWAgV_Sa

13

US patents are amazing—throwback to time when Sony really filed one for shouting brand name to skip ads away. And then there’s #India. Here, you could spend a year building a genuinely complex AI product — multi-layered models, novel training workflows, custom inference logic — and the patent office will still ask: “But where is the hardware?” It’s wild. Some of the most interesting innovations we’re creating today live completely in software. Algorithms are the product. Yet, because India still treats software as “just a computer program,” so many founders and engineers end up with ideas that are impossible to protect. I know startups that have: -> built entire ML optimization pipelines, designed intelligent orchestration systems, created proprietary AI-driven automation engines… …and were told they needed to attach it to some arbitrary device just to qualify. Feels like we’re forcing 2025 innovation into a 1990s patent framework. I’m genuinely curious — has anyone here tried patenting software or AI in India? What was your experience like? Did you have to restructure your idea to fit the hardware requirement, or just give up on protection altogether? Would love to hear real stories. It’s a bigger issue than we talk about. — #IntellectualProperty #Patents #Innovation #AIPolicy #SoftwarePatents #IndianStartups #TechRegulation #AIInnovation #IPChallenges #ProductDevelopment #DeepTech #LegalTech #StartupIndia DPIIT-IPR Chair_Centre for Study and Research in Intellectual Property Rights, NUSRL Ranchi (CSRIPR) Startup India #StartupIndia #startupindia

1

From 🇮🇳 Sarvam Vision achieves state-of-the-art accuracy of 84.3% on the olmOCR-Bench (English only subset) outperforming frontier models like Gemini 3 Pro and recent OCR models like DeepSeek OCR 2. IndiaAI

59

1 Comment

The AR, VR, and XR ecosystem is still evolving, without a clearly defined structure or consolidated community. In this discussion, Vivek Shah a, Founder of Capermint Technologies, highlights a key reality: while immersive technologies are advancing rapidly, the industry standards, scalable frameworks, and mature ecosystems are still under development. This presents both uncertainty and opportunity. Industries in early structural phases often reward those who contribute to shaping their foundations. A timely insight for founders, developers, and technology strategists exploring immersive platforms. #ARVRXR #ImmersiveTechnology #TechInnovation #GamingIndustry #FutureOfTech #LeadershipInsights

14

RIP Robotic text — I found a Humanise AI tool that makes everything 10x more natural! Introducing ChatLLM by Abacus.AI . All AI tools in one place! - http://chatllm.abacus.ai HERE ARE ALL THE FANTASTIC FEATURES 1. ACCESS TO ALL THE SOTA LLMs ChatGPT- 5 Pro, o4, Claude 4, Gemini 2.5, Grok 4.1, DeepSeek, and many more 2. IMAGE & VIDEO GENERATION Using the best AI models, FLUX, DALL-E, Ideogram, Kling, Runway, Luma & many other tools to turn your ideas into amazing pictures & videos. 3. BUILD APPS & WEBSITES A simple prompt instantly gives you live website previews. Edit everything—colors, text, layout—effortlessly! 4. AI PRESENTATIONS ChatLLM allows you to generate Word, Excel, or PowerPoint files quickly with a single prompt. It is quick, simple, and time-saving. 5. AI PERSONAL ASSISTANT - OPERATOR AI handles tasks for you 24/7, even while you’re asleep. It can search the web and take action, acting like a personal assistant that never needs a break. 6. CHAT WITH PDF You can work with any document, no matter how long or short & when things get complicated, get a quick summary that makes sense. 7. HUMANIZE AI CONTENT RIP AI Detectors. Generate your AI content. Just click the "Humanise" icon, choose your preferred tone, and the text will become human. 8. SEARCH THE WEB RIP Perplexity - ChatLLM gives real-time answers with trusted sources, images, and videos. 9. DEEPAGENT (To literally do everything) Create mobile Apps, Websites, Technical reports, presentations and much more with simple prompts. Make edits with simple english! 10. AI CODE EDITOR + ASSISTANT: CODELLM This AI Code Editor helps you write, run, and perfect your code with ease. Vibe Coding it is! This all-in-one tool can replace all your AI tools for just $10 a month! Give it a shot! http://chatllm.abacus.ai

455

133 Comments

India just switched on 38,000 sovereign GPUs at about ₹65 per GPU hour. It is a bold move backed by a ₹10,300 crore commitment. Yet more than 100 Indian AI founders are reportedly relocating to the United States. #startupindia The IndiaAI Mission is now live with high end compute on sovereign rails and a clear goal. Make AI in India and make AI work for India. But the talent math is tight. Hiring data points to a shortfall of roughly 1 to 1.4 million AI professionals this year alone. Only about 16% of IT workers have skills. Around 51% of AI or ML roles are unfilled. In some generative AI roles there is one qualified engineer for every ten openings. At the same time founders are following gravity. Frontier research, dense talent clusters, early adopters, and large enterprise buyers are still concentrated in San Francisco and Seattle. That makes fundraising, hiring, and closing six and seven figure contracts easier. UPI and Aadhar AadhaarStack (National Payments Corporation Of India (NPCI)) showed India can build sovereign digital rails that the world later studies. The paradox now is we are laying AI rails at home while many of the highest leverage companies are registering abroad. If you are a young Indian AI founder the choice is stark. Build on affordable sovereign compute close to Indian users, or build near models, money, and markets overseas. What would it take for India to make that choice unnecessary within three years E2E Cloud L&T Construction CMS Computers India Pvt Ltd. Jio Platforms Limited (JPL) Locuz Enterprise Solution Limited. CtrlS Datacenters Tata Communications Yotta Data Services Private Limited

39

7 Comments

We just published at the ES-Reasoning workshop at ICLR 2026! Me, Yahia Alqurnawi, Preetom Kumar Biswas and Tejas Anvekar worked on ViTaB-A to better understand the structured data citation capabilities of current mLLMs. We kind of realized early that QA and Citation are two very separate tasks and models aren’t very performant or optimized for the latter. Here are some other results we concluded in the paper: 1. Surprisingly, citations in images are easier than MD and JSON. This isn’t the case for QA though. 2. Citing rows is easier than citing columns. 3. Confidence scores (internal or verbal) are not a good verification mechanism for how the model might have performed. Also, a huge thank you to Dr. Chitta Baral and Dr. Vivek Gupta for all of their support and guidance on this. If you’re interested in reading more, here’s the preprint: https://lnkd.in/gmTDV__5 We’ll keep updating the paper with any new insights or numbers we find. School of Computing and Augmented Intelligence #research #interpretability #nlp

34

10 Comments

First TTS model from Ringg Labs focused on Indian English and Hindi. Contact admin@ringg.ai for TTS API access, on-prem deployments and voice cloning. More models in the pipeline which run on CPU and support for other Indian languages. Made in India for India. If we get enough interest, we'll release an open source model too :)

6

##My Software Tool – 93: 🌍 Introducing “VakyaVadanX”: India’s First Auto-Updating Narration Tool. I’m excited to unveil VakyaVadanX, a next-generation, browser-based Narration Solver that transforms direct and indirect speech into a live, self-updating grammatical system — no Convert button, no Run action, no manual refresh. Just start typing a sentence and select the narration mode → the converted narration updates instantly, strictly following exam-standard grammar rules. 🎥 Demo video attached to showcase the auto-updating narration workflow in action. 🔍 What VakyaVadanX Analyzes & Converts (Live & Automatically) 🔹 Direct → Indirect Speech 🔹 Indirect → Direct Speech 🔹 Automatic tense back-shift and restoration 🔹 Time & place word transformation (now → then, today → that day, here → there) 🔹 Pronoun normalization 🔹 Proper capitalization and punctuation as per grammar rules Each output updates continuously and consistently as the sentence or narration mode changes. 🧠 Why VakyaVadanX is Different 🔄 Fully Auto-Updating Grammar Engine (type/select → narration updates instantly) ✔ No Compute / Run / Convert buttons ✔ Rule-based, transparent narration logic ✔ Correct handling of tense back-shift & restoration ✔ Proper uppercase & lowercase control for exam accuracy ✔ Ideal for school grammar, competitive exams & foundational NLP learning ✔ Classroom-ready, exam-ready & research-oriented ✔ No Python. No NLP libraries. No APIs. Fully Browser-Based VakyaVadanX is designed for students, educators, and language learners who want to understand narration as a living grammatical process, not just memorized rules. 📘 Copyright: VakyaVadanX © 2025 🛠️ Designed & Developed by: Ankit Kishore Sen #GrammarTech #EdTechIndia #NarrationSolver #EnglishGrammar #LanguageLearning #NLPFoundations #MadeInIndia #Innovation #GrammarVisualization

3

My last post on Kompact AI unexpectedly went viral (1.5M+ impressions on April 12th) and sparked a wave of diverse perspectives — especially around how India might fall behind in the LLM and GPU race. But if we rewind to 2010, the world was racing toward digitisation — and India seemed far behind. We had: • No landlines in most homes • No widespread credit card usage • Few payment machines • High internet costs • And let’s be honest — most people didn’t even know how to send an SMS A familiar pattern emerges. Every tech revolution came with the same set of challenges for India: Infrastructure gaps. Capability gaps. Cost barriers. Scale problems. And yet, every single time, India didn’t just catch up — it leapfrogged. • We skipped landlines and went straight to mobile • Skipped credit cards and jumped straight to UPI • Created Aadhaar, the world’s largest digital identity system • Went from vaccine helplessness to building a global benchmark with CoWIN So why not AI? While the world innovates for efficiency, India innovates to survive. While the world builds for the top 10%, India builds for the last person in the queue. We may not play the same game, but we play it in our own Jugaadu style — and often, we end up redefining the rules. AI in India won’t just be about chasing GPUs and LLMs. It will be about reimagining how technology can solve grassroots problems at scale. And when that happens — it’s not just catch-up. It’s transformation. #India #AI #DeepTech #UPI #Innovation #KompactAI #Jugaad #TechForGood #DigitalIndia

115

18 Comments

Visual Pre‑Training on Unlabeled Images using Reinforcement Learning by Sergey Levine and Dibya Ghosh. TL;DR - Combines self-supervised pretraining with RL-style value learning on unlabeled images. arXiv - Frames augmentations and view changes as actions in a “dynamics” over images. - Learns a general value function to associate views with expected future returns. - Analogous to crop-consistency SSL—but flexible via reward shaping. - Trained on web-scale unlabeled images (EpicKitchens, COCO, CC12M). arXiv - Achieves stronger representations for downstream tasks versus standard SSL. In brief: Visual Pre-Training on Unlabeled Images using Reinforcement Learning” reframes SSL as an RL task: treat augmentation as transitions and train a value function predicting future consistency. It learns features that better capture decision-relevant structure. Experiments on COCO, EpicKitchens, CC12M show improved downstream performance. It’s a simple, scalable idea that combines the best of SSL and RL. Some of the limitations include potential mismatch between augmentation dynamics and real-world tasks; future work could combine reward-based pretraining with contrastive learning, apply to multimodal data, or explore curriculum shaping of transformations. A very interesting for RL and Multimodal enthusiasts. https://lnkd.in/gwskHz4K #calberkeley #bair #ai #rl #multimodal #academic #research #idea #stanford #openai #anthropic #agi #llm #representationlearning #reinforcementlearning #umass #berkeley #bayarea #boston #austin #uta #alberta #banglore #delhi #zurich #ethzurich #reinforcementlearning

4

CHINGMU Vision made its debut and concluded successfully at IROS 2025. How does motion capture deeply empower the field of embodied AI? What profound insights do more than a thousand experts and scholars share? Let's look back on this grand event in the field of embodied AI together!#IROS #robot #mocap #ai

2

We just tested Sarvam AI's Text-to-Speech and Speech-to-Text models with our AI agent — here's what we learned. India's AI ecosystem is maturing fast. But do India-focused models actually outperform global alternatives for local use cases? We decided to find out. **What we tested:** ✅ Integration into our existing voice agent ✅ Hindi conversation quality ✅ Code-switching (Hindi-English) ✅ Latency and reliability **What we found:** Some wins, some surprises, and a few gotchas that'll save you debugging time if you're building with Indian AI models. This isn't a marketing piece — it's 14 minutes of honest technical breakdown with live demos, API integration walkthrough, and side-by-side comparisons. 👉 Watch here: https://lnkd.in/gCibuxNu Whether you're building AI products for India or evaluating local vs global AI solutions, this will give you real-world context. #AI #VoiceAI #SarvamAI #IndianAI #TechInIndia #AIEngineering #ProductDevelopment

28

3 Comments

🧠 𝐓𝐡𝐞 𝐦𝐨𝐦𝐞𝐧𝐭 𝐀𝐈 𝐬𝐮𝐫𝐩𝐚𝐬𝐬𝐞𝐝 𝐡𝐮𝐦𝐚𝐧 𝐏𝐡𝐃𝐬 𝐨𝐧 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 𝐛𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐬 OpenAI's o3-Pro just achieved something remarkable – outperforming human PhDs on intelligence benchmarks by significant margins. The ripple effects are already transforming how we work. Here's what I'm seeing firsthand: My co-founder Pranav has turned o3-Pro into his research thought partner. Instead of spending weeks combing through literature, he asks o3-Pro to analyze hundreds of papers and surface the most promising algorithmic approaches... within just 15 minutes. When he finds something compelling, o3-Pro pulls the relevant code from GitHub and integrates it for testing – cutting his evaluation time by 5x or more. But the real breakthrough? Democratization. Team members who previously couldn't tackle complex research are now producing work that rivals seasoned researchers. I'm seeing a clear divide emerge in organizations. 💡 The people who are curious and love what they do learn more and tend to "play" more. They are bringing in AI-transformation in our company. 🥣 The traditional manager who focuses on people? The project manager who focuses on co-ordination? The engineer who does the tasks they are given and then goes home? These types are often unable to bring in AI-transformation in organizations, I'm finding... with those sorts of people, often I have to spoon-feed them AI approaches. Some of them take the spoon-feeding and deliver. Others hastily evaluate something and come up with some reasons why AI doesn't help them that much and continue business-as-usual. The cost of business-as-usual can often be 3x more time or 2x-5x higher cost or significantly lower quality or a combination of these. At least that's what I'm finding in our company. I've seen enough of AI and its trajectory to say this: Stay hungry, stay (AI) curious. It is going to change the world.

32

4 Comments

India’s AI ecosystem is entering a decisive decade. When leaders shaping foundational AI systems engage with researchers, industry, and policy stakeholders in one room — the conversation changes. Research Heights Foundation (RHF) honored to share that Prof. Ganesh Ramakrishnan Lead & Principal Investigator – BharatGenAI, will be part of the National Conclave on Research & Deep Technologies for Sustainable Bharat 2047. 📅 28 Feb – 1 Mar 2026 📍 Hyderabad This conclave is focused on one serious question: How do we build sovereign, scalable, production-ready deep technologies for Bharat? From applied AI and digital engineering to IP creation and research commercialization, the discussions are designed for decision-makers — not spectators. If you are working on: • Applied AI & large-scale systems • Research translation & industry deployment • National capability building in deep tech You should consider being in the room. 🔗 Registrations open: https://lnkd.in/grt9rrpK The next phase of India’s innovation journey will not be built in isolation. It will be built through convergence. #DeepTechIndia #AIForBharat #ResearchToImpact #NationalScienceDay2026 #SustainableBharat2047 #RHF #ResearchHeightsFoundation Krishnakumari Palle (KK) Amjad khan Patan Dr.Durga Prakash Devarakonda (DP) AMARNATH JULURI Dr.Dinesh Chandrasekar DC Prithvi Chaithanya Kotha Naga Jyothi Avusingi Abhinavan Sarikonda ✨ Amith Singhee Atul Mangal Kavita Bhatia Dr. JBV Reddy Sridhar C V

14

Recently, I got the opportunity to be on Vishal Malhotra’s podcast to talk about AI. This wasn’t a basic “AI is the future” conversation. We discussed thing like: AI voice cloning scams AI leading to war AI replacing actors Students over-depending on tools where India stands in the AI race and even whether AI could develop self-preservation instincts you won’t just learn how you can integrate AI into your life. You’ll understand what it means for media, creativity, education, and human thinking. Because AI is not just a tech upgrade anymore. It’s a shift in how we live, create, and make decisions. Watch it now Link: https://lnkd.in/dB3dgtyp

1

I love seeing ML teams who refuse to treat biology as an afterthought. Relation ‘s PatchDNA work is one of the clearest examples this year of what happens when you start with how biology actually works and only then design the model. No shortcuts, no flattening of the problem just to fit the tooling. That mindset hits close to home for me. At Outpost Bio, we’re obsessed with the same question: how do you build AI that doesn’t attempt to overwrite biology, but explains it? When you’re dealing with trillions of microbes shaping what we absorb, how drugs behave, and how our bodies respond to the world (biology’s “hidden majority” that most research ignores), you can’t get away with convenient abstractions. You have to let the system speak for itself. PatchDNA is a great example of doing that well: a biologically grounded inductive bias for DNA that gives the model structure instead of asking it to guess. Huge congrats to the Relation + Rosalind teams on the NeurIPS Best Paper. This is the direction the field needs: ML that respects mechanism. Congratulations! Cristian Regep Alice Del Vecchio Lindsay Edwards Rosie Rodriguez PhD Jake P. Taylor-King Sara-Jane Dunn #NeurIPS2025 #NeurIPS 🏆 Best Paper Award: NeurIPS AI4D3 Workshop PatchDNA: A flexible and biologically informed alternative to tokenization for DNA https://lnkd.in/dPHswTGP 🎤 Talk: Saturday, December 6th, 11:30 Convention Center, Upper Level, Room 28A–E 🖼️ Poster Session: Sunday, December 7th, 15:55–17:00 AI4Science Workshop, Convention Center, Upper Level, Ballroom 20A

85

4 Comments