Ravi Choudhary

Fractal, a Mumbai-based AI firm, launched Fathom-R1-14B, a 14B-parameter open-source language model excelling in math reasoning. It outperforms o1-mini and o3-mini, nearing o4-mini levels, costing $499 post-training. Available on Hugging Face and GitHub under MIT license, it was built under IndiaAI’s mission and scores up to 76.7% and 56.7% on AIME-25 and HMMT-25 exams with extra compute. Fine-tuned with supervised learning and model merging, a variant using reinforcement learning costs $967. Fractal also created Vaidya.ai, a free healthcare AI platform, while Sarvam launched Sarvam-M, a 24B-parameter model under IndiaAI.

401

6 Comments

Indian Industries should hire Pure Mathematicians So far the only group of Industries I have seen proactively willing to hire pure mathematicians are either hedge fund / investment banks / Casino-Gaming companies or AI companies who wants trainers for their AI. It is somewhat humbling and mildly humiliating trying to convince recruiters that other than the aforementioned roles and teaching there are other roles where people with pure mathematics background can contribute and it would be worthwhile for both parties. Question is Why, Where and How ? Some generic reasons which I can think of 1) A Pure Mathematician can think. Think on their own. 2) They are trained Researchers. Can read and learn from the very basic to very advanced level from technical document. 3) Able to abstract lot of information in short / precise statements ... no fluff 4) Able to listen and understand and communicate with almost any branch of Engineering and Sciences 5) Have got extensive peer network. 6) Can both work individually as well as collaborate ... When you are hiring such a person the fellow has 8-10 years of research experience, extensive knowledge base, problem solving skill, determination and inclination to work on tough problems which are open ended. PS: After reading 5 Graduate Level Textbooks and 20 research paper the fellow will introduce himself/herself as a beginner in his field.

451

57 Comments

Recently, PW (Physics Wallah) launched Aryabhata 1.0, an AI model built using something called an SLM, which stands for Small Language Model. Unlike big models like GPT-4 or Gemini (called LLMs – Large Language Models) that need a lot of data, power, and memory to run, SLMs are smaller, faster, and easier to use. They are made for specific tasks and give quick answers without using too many resources. This makes them a good fit for education, where speed and accuracy are more important than handling every type of question. While LLMs try to do everything, SLMs are focused — perfect for things like doubt-solving, tutoring, and exam help. Note: A lot of news has surfaced claiming that it is just a wrapper on top of the Qwen model. However,this post is solely focused on how LLM and SLM models differ.

730

8 Comments

🚀 Initiative to build an Open-Source LLM for Konkani Konkani, an official language of India, has a rich heritage, diverse dialects, and it is written in multiple scripts across different regions: ✅ Devanagari – Officially recognized in Goa, widely used in literature and formal settings. ✅ Roman (Romi Konkani) – A legacy of Portuguese rule, still dominant in media, Tiatr, and religious writings. ✅ Kannada – Used in Karnataka, particularly in Uttara Kannada. ✅ Malayalam (Obsolete) – Historically used in Kerala, but most Konkani speakers there now use Devanagari. ✅ Perso-Arabic (Obsolete) – Once used by Konkani-speaking Muslims, incorporating Urdu & Arabic influences. While Devanagari holds official status, limiting AI models to a single script ignores a significant portion of Konkani speakers. Many native speakers read, write, and communicate in different scripts, making it crucial to incorporate all of them when building effective AI models. Konkani’s digital presence is extremely limited, and most existing datasets are proprietary. By leveraging all available scripts, we can enhance AI models, even if the scripts differ, their underlying linguistic structure is closely connected. A script-inclusive approach strengthens the model’s ability to generalize across variations, and should benefit different NLP applications. 💡 The Initiative Launching an open-source effort to build a high-quality Konkani corpus across all scripts; no restrictions, no exclusions. 📌 The first version of the dataset, extracted from Wikipedia, is now available on Hugging Face. However, this is just the beginning! The current dataset is very limited, and we need a much larger, high-quality corpus to build strong AI models for Konkani. 📢 Call to Action If you’re a researcher, student, media outlet, or language authority, we invite you to contribute data, resources, and expertise. Without a strong digital presence, Konkani risks fading in the AI era. Let’s change that together. Codebase & Dataset Repo GitHub: https://lnkd.in/dw4gwBu2 Hugging Face: https://lnkd.in/daaKgxDG This initiative requires collaboration from various stakeholders—researchers, government authorities, academicians, language experts, and tech enthusiasts. Let’s build a state-of-the-art Konkani LLM together! #Konkani #AI #LLM #NLP #OpenSource #HuggingFace #MachineLearning #ArtificialIntelligence #GenerativeAI Government of Goa Goa Konkani Akademi Goa University World Konkani Centre

107

18 Comments

Proud to Announce: BharatGen - Param 1 | India’s Bilingual Foundation AI Model #AIforBharat #IndicAI #DigitalIndia Introducing Param 1, a 2.9-billion parameter bilingual foundation model trained on high-quality, culturally rich datasets in English and Hindi — powering the next generation of Indic AI systems. Developed under the BharatGen initiative by TIH Foundation for IoT & IoE, Param 1 is optimized for text generation and completion, built on approximately 5 trillion tokens of domain-diverse Indian data. It delivers state-of-the-art performance on bilingual tasks while maintaining computational efficiency. Key Highlights: • English-Hindi bilingual foundation • 2.9B parameters • Trained on 5T+ culturally grounded tokens • MIT licensed and openly accessible • Hosted on AIKosh, IndiaAI’s official platform This is a proud step towards India’s vision of sovereign, inclusive, and accessible AI. Explore more and access the model here: https://lnkd.in/dHcBAD66 Developed by: TIH Foundation for IoT & IoE | BharatGen Team #ArtificialIntelligence #LanguageModel #OpenSourceAI #IndiaAI #TechForGood #BharatGPT

40

What can one outlier do to Correlation ? Ans: A lot Correlation is Fragile. One Outlier can change it !! Yesterday, Venkat Raman shared a really pertinent point: MMM ≠ Correlation. ICYMI, link to the post in comments. Some discredit MMM by falsely calling it correlation based. However, in the same breath they heavily rely on correlation themselves !! Over reliance on correlation is dangerous. I see countless posts and even research papers where marketers confidently assert that some marketing/media variable is "highly correlated" with a KPI - implying causation without questioning the metric itself. But here’s the problem: correlation is incredibly prone to outliers among other flaws. In this illustration, we show how adding just one outlier can dramatically change the correlation coefficient. Yes, just one point. That’s how non robust correlation can be as a standalone metric. Yet, many marketers put blind faith in it. But here’s the thing - just like we have tools beyond a hammer, we have far better tools than correlation. Better correlation metrics than the usual Pearson correlation coefficient themselves. At our workshop we emphasize advance correlation techniques like Spearman correlation and Kendall's tau as well. Moving beyond the world of correlation, we do have one of the most powerful tools out there -> Multi-linear regression. Unlike simple correlation, multi-linear regression helps account for multiple variables simultaneously, making it far more robust. It powers Marketing Mix Models (MMM) for attribution, aids in causality analysis, and provides far deeper insights than a single correlation coefficient ever could. 📌 So why do so many still cling to correlation? Because it is easy. It gives a number, and numbers feel like proof. But in reality, correlation is often just an illusion of confidence - a metric that can mislead when taken at face value. Next time you see a post celebrating correlation, ask: ✅ Is it robust? ✅ Is it stable against outliers? ✅ Does it actually mean anything for decision making? Because without these checks, one is just chasing numbers, not insights. And remember, correlation could be a hint for causation but not causation itself.

58

9 Comments

Super excited for Sarvam-M to be opened up to the world! A 24B hybrid model that reads India’s languages, solves maths, and writes code with ease. 🌟 Why we’re excited - Made in India, for India – average +20 % uplift on Indic-language benchmarks, with a massive +86 % jump on romanised GSM-8K. - Brains and speed – a single model with switchable “think / non-think” modes for chain-of-thought depth or fast replies. - Punches above its weight – outperforms Llama-4 Scout and rivals 70 B-class models like Llama-3.3 on most tasks. - Math & code wizardry – +21.6 % on maths, +17.6 % on programming; HumanEval 88 % pass rate! - Lightning inference – FP8 post-training quantization + look-ahead decoding deliver up to 2× throughput on H100 GPUs. Sarvam is building an end-to-end AI ecosystem for India, empowering consumers, businesses, government bodies, and universities to plug cutting-edge models straight into their workflows. Get started with the open weights on HuggingFace. Detailed technical blog: https://lnkd.in/gpCwWSgJ #SarvamM #OpenSource #SovereignAI #IndicAI #LLM

239

5 Comments

From spending patterns to search trends, signals are everywhere. Traditional research may not be enough to capture them all - that’s where Systematic Active Equity (SAE)* steps in. By combining data, AI, and human insight, it transforms scattered signals, aiming to build sharper foresight. JioBlackRock Flexi Cap Fund. NFO period: 23 Sep - 7 Oct 2025. *Systematic Active Equity: A technology-driven** investment approach that combines human expertise with a data-driven model to build a portfolio. **The systematic approach involves utilizing inputs from the Fund Managers and signal research scores shared by BlackRock Inc. #JioBlackRockMutualFund #SystematicInvesting #WeAreInvestedInYou

66

Fathom-R1-14B, developed by the AI Research team at Fractal, was featured in Mint in a recent roundup of India's open-source AI models - recognized as one of the most downloaded among homegrown releases in May 2025.   Fathom is a powerful 14B parameter reasoning model designed for accuracy, transparency, and adaptability. It can also support a wide range of enterprise-grade applications, while enabling researchers and developers to build responsibly on top of it.   Still early days, but encouraging to see momentum building for open, reasoning-focused models developed in India.   Model: https://lnkd.in/dF6Qxs-d

179

3 Comments

18,693 GPUs to build AI for India! Substantial subsidies from the IndiaAI Mission are here to boost our AI economy. But, this warrants a question — are we subsidizing the right outcomes, or should we realign our scopes? 🔵 CRUX: Subsidies are necessary, but even policymakers agree that private capital inflows are critical to finding the next OpenAI, or building a new Nvidia for that matter. This begs a very important question: who truly benefits from government programs on AI? 1️⃣ INDIA OR THE WORLD?: Subsidies often have unintended consequences. Unemployment benefits in various economies have proved that they disincentivize work, than uplifting society. Drawing a parallel with India’s GPU incentives, this could encourage our startups to use subsidized AI infrastructure to build AI solutions for the world. 2️⃣ THE SERVICES ROUTE: For natural reasons, most startups follow economic incentives to channel the money they’ve raised into products with clear RoIs. The highest-paying customers for AI today are Western enterprises. Left to market forces, will Indian AI startups end up building for the West — much like India’s gigantic (and reasonably successful) IT services industry? This can have negative consequences — low margins, and a deep reliance on global markets where success gets linked directly to variables not in our control. This is far from ideal. 3️⃣ A FUNDAMENTAL BOOST: Instead of the low-hanging fruit, as an investor bullish on India’s AI story, I’m keen on startups building fundamentally to solve India’s unique issues. Take healthcare and education, for example: while GPU subsidies are necessary, India should invest more on AI tutors for students, or AI doctors for rural healthcare. In agriculture, the IndiaAI Mission should give bigger boosts to AI for farmers — as part of its proposed marketplace model. This can give actual, on-ground benefit through AI interventions in conventional farms. 4️⃣ LONG-TERM THINKING: Using GPUs to build apps for the world will increase supply in a crowding AI market. But, building impactful AI solutions will create demand for AI in everyday living — showing actual results that users can see. This is of far greater value — both empirically and relatively — than just catering to the world. 5️⃣ UPI FOR AI?: We must look at what UPI has achieved. IndiaAI can create public AI models that make all of the above possible — AI tutors, farm assistants, and remote medical help. Startups can innovate with local datasets to build applications atop our own foundational model. This model is attractive to us investors: SIDBI’s Fund of Funds has already proved that. Let’s replicate it for AI, too! Think of this as the early 2010s, and we’re building an underlying framework that would one day give birth to UPI. Sounds exciting? Should we focus more on made-for-India applications, and not just compute? Join me in comments!

111

14 Comments

When will the chase towards building for optics stop within Indian institutions? When will the obsession with optics stop? Especially in Indian institutions. Every demo is polished. Every initiative is headline-ready. Every deck is photo-op perfect. But scratch the surface, and there’s… → No adoption. → No ownership. → No actual user. I’ve sat in rooms where “AI pilots” were proudly shown off — built by third-party vendors, approved by three committees, celebrated on LinkedIn… but not used even once by the actual team it was built for. This isn’t innovation. It’s institutional theater. If we want India to truly lead the AI wave — we’ve got to stop building for press releases. And start building for people. Adoption > Applause Outcomes > Optics Usefulness > Vanity

33

Why the World Is Underrating Indian AI - But Fractal, Haptik & Others Are Changing That Most people think Indian AI is “still catching up.” They couldn’t be more wrong. Let me tell you why... Fractal Analytics just crossed ($311 million) ₹2,765 crore in revenue with a $2.44 billion valuation. This isn’t a small player; it’s a heavyweight proving India can build AI powerhouses that compete globally. Haptik grew 68% and got acquired by Jio for ₹700 crores. Think about that — an Indian AI startup becoming a core asset for India’s largest digital conglomerate. InVideo is quietly crushing it, hitting $70 million ARR by leading AI video generation innovation. The kind of tech mostly seen in Silicon Valley—built from India. Mad Street Den pulled in $28 million with AI-powered retail solutions transforming brands and customer experiences. And then there’s AtomicWork, Sarvam AI, and a growing list of disruptors who refuse to be underestimated. Here’s the truth: Indian AI isn’t a small footnote in the global story - it’s a rising chapter reshaping industries and creating value at scale. These companies are building solutions born from Indian complexity, solving problems no one else can, and scaling faster than most expect. The world underrates us now… but that’s changing fast. Keep watching this space; the true surge of Indian AI is just getting started. #AI #India #GrowthMindset

133

8 Comments

𝐇𝐨𝐧𝐨𝐫𝐞𝐝 𝐭𝐨 𝐛𝐞 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐝 𝐢𝐧 The Times Of India The rise of models like DeepSeek AI has sparked a big debate—can knowledge distillation help India build its own large language models (LLMs) faster? While this technique offers speed and efficiency, is it enough to build truly Indian AI? My take? Optimizing global models can only take us so far. If India wants to lead, we need AI that isn’t just trained on our languages but deeply understands our economy, industries, and strategic needs. AI should be more than a technological race—it should drive India’s growth, autonomy, and competitive edge. If we want AI that truly represents India, we need to 𝐛𝐮𝐢𝐥𝐝 𝐟𝐫𝐨𝐦 𝐟𝐢𝐫𝐬𝐭 𝐩𝐫𝐢𝐧𝐜𝐢𝐩𝐥𝐞𝐬, not borrowed intelligence. This article explores how India should navigate this AI shift, with insights from industry leaders. Read more here: https://lnkd.in/ghZ5B5iX #Enlog #Deepseek #AI #TechforGood #EnergyManagement # Smartsolutions #TOITechies #TOI Enlog Bharath Rnkawat Jharna Saha

42

3 Comments

How can I ride the AI wave through public markets in India? I asked this research question on ChatGPT 5Pro: "Which industries and subsequently type of companies will benefit substantial from the introduction of AI in mainstream business? Research recent past results of public companies, earnings calls of ceos and cfos and discussions to provide insights What’s the purpose of this report? As a public market investor in India, I want to get ahead of the curve and invest in publicly listed companies (from any industry) that will benefit from AI in their earnings over the next few quarters, thereby providing an edge for investing in them before the market catches on." The answer is available to read as a PDF in the link below! (Gimme feedback if this helps you and you want more! Also, if you wanna partner with me on researching and investing in Indian/Global markets, just ping or write me at alok@Games2win-dot-com #Dhandhekibaat https://lnkd.in/dp47ZCVZ

19

7 Comments