About
I’m a 22-year product & engineering leader who builds high-trust teams to ship…
Activity
2K followers
-
Saurabh Vajpayee posted thisEra of AI Labs: Technology has gone through a lot of change in last 40 years. From Desktop PC, to Internet, to Mobile Phone, to Social Media, to Cloud Computing and now is the Era of AI Labs. And I reckon that these AI Labs will become dominant and most important companies in near and upcoming future. The power of an AI Lab: AI Lab is the place where the meat or actual meal (if you are a vegan) of the AI revolution is produced. Everything else - is a support function, which I believe the AI Labs will manage in house soon. It is here, that MLPs models - which we call - LLMs and more - are coded, trained, security checked and released. The chip makers, the data centers and the cloud partners - though bigger now - will become much smaller as these labs produce smarter AIs and grow dominant and want to vertically integrate. Unlike social media - where social media is owned by 1 company, the phone by multiple other companies - an AI Lab can venture into Chip Design, Data Center Ops, Cloud Layer - when it chooses to. The power of an AI Lab comes squarely from the models. These models can reason, code, write documents, emails, create videos, pictures and increasingly go into agentic commerce. Thus, an AI Lab can potentially disrupt a much wider industry - from IT Services, BPOs, SaaS, Banking, Payments, eCommerce, Education - pretty much all segment of the economy. A rogue AI Lab can even harm geo politics, defence and cyber security. At present, the swings in sentiment and markets is superficial, but soon, these will be for real. Markets globally are trying to adjust for this "AI Labs" dominant world - predict and future price. Is AI Lab only a USA and China phenomenon? No. We know about USA labs (dominated by OpenAI and Anthropic) and China (DeepSeek, Qwen by Alibaba, Moonshot AI etc). Labs are also there in Europe (Mistral, DeepMind, Stability AI). So it is a global phenomenon now. How about India? Do we have AI Labs? Yes, we do. Your Sarvam.AI is an AI Lab. But not much else. There are labs buried inside research and academic institutes - but those cannot scale and become big like a private company. So there is definitely scope of more AI Labs in India - a country of 1.5 billion people. What is preventing AI Labs to take hold in India? Two factors: 1. Access to capital. Our whole risk capital model is build for either labour arbitrage or product arbitrage. So for our investors, an AI Lab is a new beast. They havent understood its importance and how to value it. 2. Protection from Foreign Labs: This is not provided by Government of India. So India's AI Labs will remain mostly below Frontier Models. But still, they have space to operate below Frontier Models. I think - time has come for Government (which I think now gets it), Investors (which are still accessing and aligning bets - so early days for them) and people (which have no clue) - what is the future - a future dominated by AI Labs.
-
Saurabh Vajpayee posted thisDear DeepTech VCs and Angels: Just wanted to share an update. Our SarvBol model has now learnt a lot of Hindi and its dialects and is handsomely beating competition on accuracy and model unit economics. This is not a theoretical data - this is actual data and comparison on official benchmarks of Government of India and other International Benchmarking datasets. We are on track to be the top of the leaderboard in Hindi by April and have lined up to cookie cut Bengali, Telegu, Marathi, Tamil with same recipe. And in March - we will start to offer our Hindi Model - at a price point that is sure going to put pressure on competition. Our whole philosophy was two fold - make the model small, keep or beat the best in market in accuracy. And now, we are seeing that in the millions of simulations that we are running. On track to deliver our promise to the Board and to give Indians a fighting chance to be heard - at a price point which will not pinch their pocket. Jai Hind.
-
Saurabh Vajpayee posted thisAre LLMs special? Lot many of you pinged me about LLMs - why can't we use them to solve everything. Theoretically - you can. But cost and efficiency wise, you should not. There is also a myth that somehow LLMs are these God Like mystical beings, almost on the verge of being self aware and everything else in modern AI is just a puny little some simple structure. The above cannot be far from truth. Modern AI - is all based on Deep Neural Networks or MLPs (multi level perceptrons). MLPs themeselves evolve as CNNs, RNNs, LSTMs and Attention (aka Transformers). Even if you are building an AI model for say "Speech to Text", "Text to Speech", "Speech to Speech (different language)", "Text to Text (different language)" - you are using almost the same building blocks. CNNs are good at images, videos. RNNs are good at sequential data such as Speech (temporal in nature) and LSTMs are good at sequential data with long term dependencies. Attentions are good at soaking in larger meaning and making it available for prediction. Overall - most modern AI models are same underneath the cover. The difference comes in how you layer them, how many connections you provide and this layering, connections gives the parameter count. So go from say ~200M AI model to ~100B AI model - the line of code will be less than a few 100s only. That's it. Just give more layers, make the network dense and you get the explosion of params. But now comes the hard part - the bigger the network, the more data it requires to stabilize its weights and biases to do meaningful work. And more data, bigger network means much bigger compute time to train it. Infact, training is also quite standard is almost all network. You shard the data, then each model instance learns on its shard and later syncs model's weight adjustments across all shards. Fundamentally, it is that simple. But here are the steps and what makes LLMs tough to train: 1. Huge Data - Somewhat of a challenge. But you can scrap it, scan it, download it from HF and other public data sets. So yes, some effort, but you can get it. 2. Model Size - Not a problem at LLMs. LLMs architecture are open secret. You can ask Claude to code it - basic LLM, it will do that in 5 mins. 3. Compute - This is challenging and costly. You will have to create a cluster with high end GPUs (either on Cloud or On Prem), that will take a data shard, run your model code, then sync correction to weights and biases. This is really, really costly process. 4. The training recipe - this is the secret sauce. How to shard, which hyper parameters to start, how to define Loss Rate, how to decay, how to checkpoint. This is what people defend as IP. The larger the model, the harder it is to converge and generalize. So more training runs, bigger compute cost, lower probability of yielding a useful model state (called checkpoint). In my work, I also found out - that making a model smaller is also a NP hard problem.
-
Saurabh Vajpayee posted thisSpeech to Text is a NOT a solved problem for India: Siri launched in 2011. Alexa in 2014. Google Assistant in 2016. We've moved on to agents, reasoning models, and whispers of AGI. So why am I building a Speech to Text company in 2026? Because for Indian languages — it is NOT solved. Not even close. Let's look at the data. The Government of India runs Bhashini — an excellent initiative that benchmarks ASR models and acts as custodian for language AI in India. Bhashini has a public leaderboard (https://lnkd.in/gS8NHnJU). Here are the Word Error Rates of the BEST available models: Hindi — 27% Bengali — 49% Gujarati — 47% Urdu — 52% Telugu — 54% Marathi — 58% Odia — 59% Punjabi — 64% Sanskrit — 64% Malayalam — 75% Tamil — 79% Read that again. The BEST model for Tamil gets 4 out of every 5 words wrong. For Hindi — our most resourced language — 1 in 4 words is wrong. These aren't from some obscure lab. These are from the Government of India's own platform, using models from AI4Bharat (which led to the birth of Sarvam AI) and others. Even the best Indian models are far from production-grade. You can only draw two conclusions: Either Bhashini isn't being taken seriously by model developers — they see no value in getting benchmarked by an independent government agency. Or the problem is real and by NOT formally benchmarking, companies keep it under wraps. I suspect both are true. Now let's talk cost. Math for a 200-seat call center — insurance tele-screening, bank KYC, or a government helpline: → 200 agents × ~5 hrs talk time/day × 22 working days → 22,000 hours of audio per month → At cheapest rate of ₹0.50/min → ₹6.6 Lakhs/month. ~₹79 Lakhs/year. Just for transcription that is 50-75% accurate for most Indian languages. That's not a line item. That's a crater in your Opex budget. For an MSME or a district-level government department — the unit economics simply don't work. So the problem is two-fold: Accuracy — far from production-grade for real Indian speech. Cost — prohibitive for organisations that need it most. But there's a deeper issue nobody is talking about. It's not the number of languages. Everyone knows India has 22 scheduled languages and 1,600+ mother tongues. The real issue is buried in a deceptively simple question: What IS an Indian language? Think about your last phone call. Your last WhatsApp voice note. Were you speaking "Hindi"? Were you speaking "Tamil"? Or something else entirely — something no model is trained for? I'll leave you with that question. And I'll share my answer — and what we're building to solve it — in my next post. Department of Telecommunications, Ministry of Communication & IT (INDIA), BHASHINI - (Digital India BHASHINI Division)
-
Saurabh Vajpayee shared thisIs big beautiful ? There is an AI craze right now, thinking that only adding a B in your AI model can give you B dollar returns. I do not subscribe to the idea. Though I relish the feeling that in trying to run this race, these companies are setup for failure - since you are competing with super AI labs, where the letter T (trillion) in funding is not a big number. I wish them success - but I know they will collapse since no protection is built for them (like no protection was ever built for social media in India). Of course, these companies have a business plan for these models. I am sure. But let me ask - buying a bigger car or a luxury car is a great option - but how much real market is there in India for it? And when there is no protection for you, price point is your only saviour - and there bigger is not better (because inference is just traversing your model from start to finish, all layers, all params and that becomes costly). At present, this is definitely boosting India's prestige - though superficially - since there is no single reputed international journal that talks about India in same league as US and China in AI research, scale and model building. It's a valiant effort, but I have a feeling, which will go to waste - once global flood gates - which are now fully open - come and awash us with their models and their data centers. So to the companies - which launched these during India AI Impact Summit - I would say - best of luck, you did give India some national pride. But national pride is not the same as successful business in long run.
-
Saurabh Vajpayee shared thisWhat is STT (speech to text)? For us, it is mystery on how simple thing such as "Speech to Text" work. Since I have personally been working heads down to provide "efficient", "accurate" and "disruptive price point" - speech to text AI models for Indian Languages, I thought, I will take this opportunity to share with you how this all works. It is rather long read - but written in a manner that it will make it very light to read. So enjoy and comment:
-
Saurabh Vajpayee shared thisHas the AI end game for Indian Software Services Industry started? Remember, markets tend to price ahead - ie what is in future.
-
Saurabh Vajpayee posted thisHow to ride the AI wave (yes, you have to ride it): If you are a student or entry level coder: 1. Just learn basics of coding. No need to go much deeper. Your coding agent will do that job for you. Make your AI coding agent as your best friend. Your AI coding agent will get you your next contract. Yes - No Job. Get used to Project outcome based pay now. Its not bad, if you and AI agent can give better than expected outcome, you can earn 2x or 3x. If you are a parent and looking for advice: Don't restrict your kid. Let her/him explore the world and develop original ideas. What you are nurturing is unrestricted thoughts and creativity - those he/she will require and AI will do the grunt work. If you are a Mid Level Manager or 20 year+ industry veteran: Be prepared to hang your boots with dignity. Become an Angel Investor. Invest some risk money in a bunch of upcoming private AI companies in your country. Talk to their founders, understand what are they building? Basically, your career is almost certain to die in next 2 - 3 years. Your main job, is to use your experience and money, to help grow future AI companies, make money and remain relevant. If you are just a pure investor: 1. Short any stocks of companies that are service providers. Anything that is not a Product Company - you can safely short and make money. 2. Short any company that claims Software as MOAT - software is not MOAT. 3. Short companies that are in real estate or providing service to places where big IT companies are based of: Seattle, Bangalore, Hyderabad etc. Buy stocks in real estate companies in and around SFO area. 4. Buy Google, Anthrophic, Nvidia - they will be clear winners.
-
Saurabh Vajpayee posted thisSoftware Coding is no longer a human job: Opus 4.6 - Tremendous Progress... On Feb 5th, this year, Anthropic released Opus 4.6, an advance reasoning coding agent. I was playing with it yesterday, and I can tell you, that it is a step function change in how the agent goes after completing it's goal. For all you non believers - that we will always need human in loop for programming - please use Opus 4.6 and then think - what is coming next !! Opus will scan your code, come up with a bunch of hypothesis, then refine those hypothesis, then test them and this refine-test-pin point - is so marvelously done, that in my case, it found a single bit shift, buried deep within a model code that resulted in accuracy error. What I really liked about Opus 4.6 is that it will not cut corners. Most coding agents try a number of ways and then simply pick the most probable or best path forward. Opus - both have much advance reasoning, but more than that, it has spirit of not giving up. In my case, the agent kept on hunting the source of the problem for almost 2 hours and tried I dont know how many hypothesis and its own tests, till it found that one line that was the issue. These are deep neural network code. It is not easy to debug and if you think that an experience data scientist would have found it in shorter time, I don't think the issue was that trivial. My casing point is this: Opus 4.6 is a step function improvement in coding agent (and I have personally used many). If this is the speed of improvement and this skill/capability of future coding agent is true, I will amend my earlier post from "Software Coding is no longer a salaried job" to "Software Coding is no longer a human job" - which I believe is now almost true.
Saurabh Vajpayee commented on a post
1w
Congratulations. Indeed a great milestone.
Saurabh Vajpayee commented on a post
3w
Well. For early stage investment, they don't. And in age of AI -> there are new business models and new economic activities that will emerge. But unfortunately our VC community has two strong likes:1. They would like a business to show some revenue numbers (missing the point that some deep tech investments take time to mature, since they are trying to fundamentally alter a landscape). Think of this way - what if early Google and OpenAI investors in Year 1 said -> show me some customer or revenue.2. Our Indian VC community is still in age of "Mobile Apps" or "Gig Economy". They are happy to shelve money on Instant House Help App or AI Apps build on top of LLMs, not knowing that those are most risky investments to make in AI. I would go as much as to say -> Indian VC community is ripe for disruption right now. New age VCs can emerge, use extensive AI, raise funds and find alignment with companies that are building more deep tech and long term platforms than mere "Apps" or "Gig" games.
-
Saurabh Vajpayee liked thisSaurabh Vajpayee liked thisहवाएँ...मौसम का रुख बदलती है दुआएँ....मुसीबत का पल बदलती है ।। और वो जो हर पल मेरे साथ होती है वही मेरा आदि-अनंत होती है ।। भारती दयाल 25-26 /3 /2026
-
Saurabh Vajpayee liked thisSaurabh Vajpayee liked thisBetter late than never! This post is a bit delayed and long overdue, but I couldn't move on without reflecting on a massive chapter of my life. My nearly 10-year journey with Expedia Group wrapped up recently, and it was an incredible ride. Looking back, I’m proud of the "architect’s touch" I contributed to several successful launches, solving complex breadth-of-travel challenges globally. Payment Orchestration, Incentives, HomeAway Globalisation, the Vrbo rebrand, Stay stack Modernisation, Travel API, and Modern Checkout are the significant milestones that shaped my journey and experience. Expedia is a place of high scale and even higher-quality people. I truly valued the culture of growth - especially the spirited, daily debates over architecture choices! To those who challenged my designs at the whiteboard: thank you for making the work (and me) better. It was a privilege to solve tough e-commerce problems alongside such a talented group. To the leaders, mentors, and peers who made this decade so meaningful - thank you for the partnership and the memories. What’s next? I’ve taken some time to pause, reflect (and even "appear" for the CBSE Class X boards! :)), and have spent my time implementing Gen AI applications. Now, I am ready to go get what’s next!
-
Saurabh Vajpayee liked thisSaurabh Vajpayee liked thisThe doors are open. The momentum is real. Our Kolkata office is up and running! Kolkata, let's get to work. 💪
-
Saurabh Vajpayee liked thisSaurabh Vajpayee liked thisCouple of months ago, I went from leading engineering teams to going hands-on and that required working a lot on AI coding agents — Claude Code, Copilot, Lovable, Replit — on real projects, daily. As someone who's built and scaled teams at tech giants, I thought I'd pick these up fast. I was wrong. The gap between using these tools and using them well is enormous. Four things changed my output quality dramatically: 𝟭. CLAUDE.md / AGENTS.md A markdown file at your project root describing your architecture, conventions, and constraints. Without it, the agent guesses. With it, it follows. The agent stopped inventing its own naming conventions overnight. Highest-ROI setup I've found. 𝟮. Task decomposition Early on I'd describe an entire feature and get something that looked done in 10 minutes. Fascinating — untill code base grew and bugs increased exponentially in generated code. Breaking features into small, testable chunks improved agent effectiveness significantly. 𝟯. Model selection I burned two days on iteration cycles before realizing I was using a fast model for a task that needed deep reasoning. The fix: reasoning models for architecture, fast models for implementation. Auto-routing isn't there yet. 𝟰. Prompt quality The most underestimated lever. 5 minutes structuring a prompt saves 30–45 minutes of back-and-forth. Every single time. The tooling is here. The craft of using it well is still emerging. #SoftwareEngineering #CodingAgents #EngineeringLeadership #DeveloperProductivity
Experience & Education
-
DeepMindSecure Technologies Private Limited
View Saurabh’s full experience
See their title, tenure and more.
or
Recommendations received
23 people have recommended Saurabh
Join now to viewOther similar profiles
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More