About
I build and scale fintech products by converting existing channel strengths into…
Articles by Pushkar
Activity
2K followers
-
Pushkar Sinha shared thisWhat I learned building a recommendation engine for a product catalogue. A few evenings this week on MovieLens, the 100K public dataset. Not for work. I'd noticed I could describe these systems a lot more fluently than I could implement them, and wanted to see what was hiding in that gap. Here's what I got wrong, in the order I got it wrong. I started with popularity, which just ranks films by how many people watched them and shows everyone the same list. I built it as a throwaway floor to beat. It scored 0.081 on hit rate — did the film I'd hidden show up in that user's top 10. I almost didn't measure it properly. It became the most useful thing I built, because everything else got compared against it. Then item-kNN. For every pair of films, count how many people watched both, normalised by how popular each one is. If you liked Fargo, surface what Fargo's audience also liked. 0.107, up 32%. This is where I lost an evening. I swept the neighbourhood size across 32 values, plotted curves, found an optimum. Felt productive. Then I checked something I should have checked first: how many films the model could score at all. For a typical user, 300 out of 1,682. The rest scored exactly zero. And for 31% of my users, the film I'd hidden was one of those zeros — not ranked badly, invisible. No setting of any parameter was going to surface it. I'd spent the evening optimising the ranking of the third of the catalogue the model could see, while two-thirds of it was dark. Then ALS, which compresses the whole matrix into 64 numbers per user and per film. 0.127. I ran it with the library defaults first, because that's what you do. Mediocre. The parameter that mattered turned out to be one that controls how loudly a watch counts against silence — the published default is 40, and on this data 1 won, with results getting steadily worse as I raised it. I looked up why. The default was designed for clicks and plays, where not-watching is genuinely ambiguous and you need to shout over the noise. A deliberate 4-star rating is already a strong signal. The default encoded an assumption about how the data was generated, and my data wasn't generated that way. I'd used that default for an hour without once asking what it was for. The two things I'll carry: Measure what your model can see before you tune how it ranks. I had the ordering backwards and it cost me the more interesting finding. And a default is somebody else's context, frozen. Usually a reasonable one. But it was chosen against a particular kind of data, and nobody hands you the assumptions with the parameter. Neither is news to people who build recommenders. They were news to me in a way reading about them never managed. If you've built these for real — did you measure catalogue reachability? I'm curious whether 31% is normal or whether I was naive. Code and results: https://lnkd.in/gTiA6MTG #RecommendationSystems #MachineLearning #ProductManagement #LearningInPublic
-
Pushkar Sinha shared thisWhat if the next big software upgrade isn’t a new UI, but replacing hard-coded if/else journeys with a state graph where an LLM decides the next best action? LangGraph made me realise that software may be moving from “code the path” to “code the guardrails, let intelligence choose the path” — which is exciting, slightly terrifying, and very expensive if you forget token economics. 😄What If We Stopped Hard-Coding Every State Transition?What If We Stopped Hard-Coding Every State Transition?
-
Pushkar Sinha shared thisA little more time. A little less commute. And a lot more building. 🙂 I'm using this transition between chapters to go deeper into Agentic AI — RAG, MCP, tool orchestration and what happens when agents can not only understand intent, but actually act on it. This article is about one idea that particularly fascinated me: what happens when businesses start exposing their capabilities to AI agents? #AgenticAI #MCP #AIEngineering #AIProductManagement #ProductStrategy #AIAgents #GenerativeAI #SystemDesign #AIArchitecture #Fintech #ProductLeadership #BuildingInPublicWhat If Every Business Exposed Its Capabilities to AI?What If Every Business Exposed Its Capabilities to AI?
-
Pushkar Sinha shared thisI used to think sentiment classification was "easy." This week taught me otherwise. As part of my first submission for the IITB Agentic AI certification, I had a deceptively simple task: get an LLM to label 300 movie reviews as positive, negative, or neutral. No training, no fine-tuning — just prompting. Anyone can nail 2 examples in a weekend demo. Getting something to actually work, reliably, across hundreds of cases — that's a different game. A few things humbled me along the way: I thought "neutral" was one thing. It's actually at least four — mixed opinions, plain descriptions, comparisons to other films, hedged praise. A fix for one quietly broke another. I kept adding "smarter" instructions to my prompt, expecting each version to do better. Instead, my accuracy went up, then down, then down again. The simplest, plainest instruction I started with ended up beating every clever version I wrote after it. I ran the same pipeline across five different models expecting GPT or Gemini to win easily. DeepSeek quietly outperformed all of them — including on the exact cases every other model got wrong. Not because it's "the best" model out there, just the best fit for this task. I even tried combining all five models into a majority vote, assuming more opinions = better answer. It actually did worse than my single best model. The lesson underneath all of this: this is the same discipline as software engineering, where most of the real work is invisible — testing, measuring, tweaking, failing, and trying again. AI engineering just moves that same rigor one level up: from code to prompts, examples, and model choice. Nothing about it is "just typing instructions into a chatbot." Grateful for a course that makes you feel this, not just read about it. 🙌 #AgenticAI #IITB #PromptEngineering #LLM #AIEngineering
-
Pushkar Sinha shared thisWhat a Weekend at The Westin Himalayas Taught Me About Building Great Products This weekend, my family and I escaped to The Westin Himalayas near Rishikesh. And I have to say—it was one of the most thoughtfully designed hospitality experiences I've had in a long time. Nestled on a hill overlooking the Ganges, surrounded by floating monsoon clouds, the property almost feels disconnected from time. Yet what impressed me most wasn't the location. It was the product thinking behind the experience. The spacious rooms. The warm welcome. The attention to small details. The excellent food (their High Tea deserves a special mention). A beautiful infinity pool overlooking the valley. A surprisingly well-designed gaming zone with bowling and indoor games that kept everyone—from kids to adults—engaged. Everything quietly worked together to make you feel relaxed. As product people, we often think customer delight comes from adding another feature. This weekend reminded me that delight is usually the result of many small things working seamlessly together. No single touchpoint made me think, "Wow!" But by the end of the stay, I found myself saying, "I definitely want to come back." That's probably the strongest compliment any product can receive. Another thing I appreciated was the value equation. The Westin Himalayas isn't inexpensive. But it felt worth every rupee. There's a subtle but important distinction. Customers don't always seek the cheapest product. They seek the product that makes them feel they made the right decision. That's where trust is built. One more pleasant surprise was accessibility. With the new Delhi–Dehradun Expressway, it's roughly a 5-hour drive from Gurgaon, making it an ideal weekend getaway without the fatigue of a full-day journey. As someone who spends most of his weekdays thinking about AI, fintech, product strategy, and organizational design, these short breaks have become increasingly valuable. Stepping away from dashboards and meetings creates space for something we rarely schedule—reflection. The biggest ideas often arrive not while working harder, but while giving your mind the freedom to wander. For me, this weekend was exactly that. A reminder that the best products aren't always the ones with the most features. They're the ones where every interaction quietly reinforces the promise the brand made. The Westin Himalayas did exactly that. I'll certainly be going back. #CustomerExperience #ProductManagement #Hospitality #ServiceDesign #CustomerDelight #ProductThinking #Leadership #WestinHimalayas #Rishikesh #Staycation
-
Pushkar Sinha shared thisIs AI creating a new discipline between product and engineering? In the final chapter of Chip Huyen's AI Engineering, the discussion on feedback in AI systems is particularly intriguing. Traditional software engineering is deterministic, while AI is not. Improving an AI system involves more than just writing better code; it requires understanding user feedback, separating signal from noise, refining prompts, adjusting context, evaluating outputs, and continuously enhancing the system. This process resembles product management more than conventional engineering. Reflecting on my recent work with an image classification pipeline, I realized that building AI products is fundamentally an optimization problem. Key considerations include: • Should I start with a smaller model? • When should I transition to a larger model? • How can I reduce latency? • How do I minimize token costs without sacrificing accuracy? These questions blur the lines between engineering and product management, pointing to a potential shift in our industry. Historically, Product Managers focused on customers, business models, and prioritization, while Engineers concentrated on algorithms, APIs, and infrastructure. Today, much of the software creation is increasingly driven by AI. As AI takes on more coding responsibilities, the source of competitive advantage shifts. Success will not solely depend on having the best prompts but on designing superior AI systems that integrate: • Customer understanding • Business objectives • Workflow orchestration • Model selection • Feedback loops • Cost optimization • Trust and explainability In essence, architecture becomes strategy. I believe that while Product Managers will not become Engineers, and vice versa, both roles will evolve. Product leaders will take ownership of customer outcomes, AI workflows, evaluation frameworks, and optimization, while Engineering leaders will focus on latency, throughput, infrastructure, and model serving. The overlap between these roles will grow, suggesting that the future may not be a battle between Product and Engineering, but rather a focus on AI Systems Thinking—where innovation is defined by intelligent systems that learn continuously from their users. How is your organization adapting? Are Product and Engineering becoming more integrated, or is a new structure for AI teams emerging? #ArtificialIntelligence #AIEngineering #ProductManagement #SystemDesign #AIArchitecture #ProductStrategy #EngineeringLeadership #AgenticAI #GenerativeAI #AIProductManagement #SystemThinking #FutureOfWork
-
Pushkar Sinha shared thisReflecting on Weeks 3 and 4 of Prof. Arpit Agarwal's Agentic AI course at IIT Bombay, I’ve realized that learning AI goes beyond just understanding transformers, attention, embeddings, or tokenization. While those concepts are important, my biggest takeaway has been that building AI products is fundamentally an optimization problem. 💡 My Biggest Realization When tasked with classifying hundreds of thousands of images, a beginner might simply send every image to the most powerful model available, get the answer, and ship it. While this approach works, it can lead to skyrocketing cloud costs. 💸 🧠 How an AI Product Architect Thinks Instead of asking, "How do I classify these images?" the focus shifts to, "How do I achieve the same business outcome with the least latency, lowest cost, and acceptable accuracy?" This change in perspective can significantly alter the architecture. For one experiment, I redesigned the pipeline as follows: ✅ Reduce image resolution before inference. ✅ Select only the most relevant images instead of processing every available image. ✅ Use a smaller, cheaper model (Claude Haiku) for first-pass classification. ✅ Ask it to return both the predicted class and a confidence score. ✅ Only if the confidence falls below 60%, escalate the request to Claude Sonnet for a second opinion. Result: ✔ Same business outcome. ✔ ~67% lower inference cost. 🚀 The Bigger Lesson This mindset extends beyond image classification. Over the past two weeks, we’ve explored: • The impact of tokenization on API costs. • The importance of context engineering over prompt engineering. • How KV caching prevents unnecessary recomputation. • The influence of prompt structure on latency and provider-side caching. • The need to select decoding parameters based on application requirements rather than personal preference. These discussions highlight that the focus should be on making AI systems economically viable, distinguishing a demo from a production system. 🔄 A Shift in Mindset Transitioning from a product background, I find this transformation exciting. I’m now thinking: ➡️ Less about models and more about systems. ➡️ Less about prompts and more about architecture. ➡️ Less about obtaining the correct answer and more about delivering it efficiently, repeatedly, and at scale. Perhaps I’m evolving from an AI enthusiast to an AI Product Architect. 💬 Over to You What optimization techniques have surprised you the most while building AI applications? I’d love to hear your real-world examples. #AI #AgenticAI #ArtificialIntelligence #AIEngineering #ProductArchitecture #ProductManagement #LLM #GenerativeAI #AIOptimization #SystemDesign #CostOptimization #EngineeringLeadership #IITBombay #LearningInPublic
-
Pushkar Sinha shared thisA fascinating experiment, Shivanshu. I completely agree that AI enables decisions to move much closer to the people solving the problem, and that's a massive shift in organizational design. Where I'm still curious is around expertise. AI dramatically expands what an individual can accomplish, but it doesn't automatically make everyone an expert in every domain. Product strategy, business strategy, engineering architecture, finance, data science, security—each requires years of judgment that AI can augment but not instantly replace. My guess is we'll still need teams, albeit much smaller and more outcome-oriented. The debate may no longer be about Frontend vs Backend vs DevOps, but about where deep specialization still creates leverage versus where AI can abstract the complexity away. What I love about Flatland is that it shifts authority from managers to problem owners. That's a powerful idea. I'll be watching with great interest to see how Cars24 balances distributed decision-making with the cultivation of deep domain expertise over the coming years. #AI #FutureOfWork #SystemThinking #OrganizationalDesign #Leadership #ProductStrategy #BusinessStrategy #AITransformation #EngineeringLeadership #GrowthSystems #Innovation #OperatingModelPushkar Sinha shared thisHad a chat with Shivanshu Makkar, Builder – Finance and Strategy, Cars24 around how the company has eliminated job titles across its 6,000-member workforce, removed variable pay, and introduced an AI-native operating model called Flatland. A few interesting takeaways: • No traditional promotions, employees take on bigger problem statements as they grow. • Variable pay has been replaced with fixed compensation and ESOP-led wealth creation. • Offer letters now encourage employees to 'hold the problem statement closely, and roles very lightly.' Read the full story: https://lnkd.in/gjfPysyKCars24 removes titles as it bets on AI-native 'Flatland' structureCars24 removes titles as it bets on AI-native 'Flatland' structure
-
Pushkar Sinha shared thisWeek 2 of the Agentic AI program brought back a memory from nearly 25 years ago. During our first semester at IIT Kanpur, one of the toughest courses was MATH101. Everyone in the class had been a topper in their respective schools. Yet, in a batch of roughly 400 students, nearly 55 received an F and had to repeat the course. I somehow managed a B. My roommate wasn't as fortunate. I still remember him sitting quietly in the hostel room and saying: "Lagta hai ki this summer, I can't go home." When our program coordinator for the Agentic AI course recently wrote to us saying that feeling challenged is completely normal, that memory came back immediately. The first few sessions have already moved from Python exercises, asyncio and Pydantic to the foundations of Machine Learning, neural networks, LLMs and agents. The course explains the stack quite well: ML learns patterns, LLMs add language and reasoning, and agents add memory, tools and the ability to act. One realization from these early sessions: An agent is not simply an LLM wrapped in a workflow. A production-grade agent must handle ambiguity, recover from failures, know when to escalate, operate within cost and latency constraints, and deliver real business outcomes. The challenge isn't building a demo. The challenge is building a system that reliably works in the real world. Learning something genuinely new can be uncomfortable, even for experienced professionals. Perhaps the most valuable skill in AI isn't intelligence. It's persistence. Back to the next lecture. #AgenticAI #ArtificialIntelligence #MachineLearning #LLM #ContinuousLearning #IITKanpur #IITBombay #ProductLeadership #AIArchitecture #GenerativeAI
Pushkar Sinha commented on a post
1w
This is great, it means 1cr+ people have the power to realize the value of NCMC card for their daily commute and spends! A real golden opportunity for monetization! Congratulations Nishant Vasani
-
Pushkar Sinha liked thisPushkar Sinha liked thisRajat Pandit has been appointed Head of Applied AI - India at Anthropic, where he will drive enterprise AI adoption across India. With over two decades of experience in AI infrastructure, cloud, and technology transformation, he previously led AI Infrastructure – APAC at Google. Earlier, he held leadership roles at Tesco and Accenture, building expertise across AI, cloud, platform engineering, and enterprise technology. #LeadershipMove #Anthropic #AppliedAI #AILeadership #kalviumx
-
Pushkar Sinha liked thisPushkar Sinha liked thisGurugram-based digital spiritual wellness platform InstaAstro has raised $12 million in Series A funding led by Singularity AMC and Artha Venture Fund, with participation from its founders. The company will use the fresh capital to expand across regional and international markets, strengthen AI-led products, and scale its Pooja Seva and spiritual commerce businesses. 🔗 Read full article link in comments Nitin Verma | Anirudh A D. #InstaAstro #DigitalSpiritual #Astrology #Funding #IndianStartups #Viestories
-
Pushkar Sinha liked thisPushkar Sinha liked thisInstaAstro raises $12 Mn in Series A round Gurugram-based digital spiritual wellness platform InstaAstro has raised $12 million in a Series A funding round led by Singularity AMC and Artha Venture Fund, with participation from the company’s founders. The round comes more than two years after InstaAstro raised Rs 18.5 crore ($2.2 million) in a pre-Series A round led by Artha Venture Fund. In November 2021, the company had raised Rs 3.2 crore in a seed round. The fresh capital will be used to expand InstaAstro’s presence in regional and international markets and invest in AI-led product development. The company plans to deploy AI for sentiment analysis and multilingual support to enhance user consultations. Founded in 2021 by Nitin Verma, InstaAstro started as an astrology consultation app and has since expanded into a broader spiritual wellness platform offering astrology, tarot, numerology, vastu, Pooja Seva and spiritual commerce. It currently has 12 million registered users and more than 5,000 verified experts across 183 countries. https://lnkd.in/dS-B9EHY
-
Pushkar Sinha liked thisPushkar Sinha liked thisThe most expensive asset in a classroom costs absolutely zero rupees. 💡 It isn't a high-tech smart board, a complex software subscription, or an expensive projector setup. It is the teacher's dedication, kindness, and vision. A great teacher can alter the trajectory of a child's life, something that no amount of capital can buy. However, even the most passionate educator cannot teach effectively when baseline logistics fall apart: 🪑 Students struggling to write while sitting on cold, dusty floors. 🖤 Faded, damaged blackboards that hinder clear lesson delivery. 💧 A lack of safe, clean drinking water that causes daily absenteeism. ⚽ Missing basic sports and learning kits that build physical and cognitive skills. Teaching is an art, but infrastructure is the foundation. When we upgrade a rural classroom, we aren't replacing the teacher; we are empowering them with the exact tools they need to unlock every child's potential. In conventional philanthropy, donors and corporate CSR teams often encounter a "trust deficit". Contributions get absorbed by administrative run-rates without clear proof of on-ground execution. At TrueGiv, we solve this through Asset-Tracking Philanthropy: ✓Ground-Truth Verification: We partner directly with verified rural government schools to log exact infrastructure needs. ✓Geotagged Asset Delivery: Your donation converts directly into tangible hardware from study desks to water purifiers, linked directly to verifiable school locations. ✓Auditable Compliance: We simplify your year-end financial planning with unified annual certificates for tax saving donations section 80G. The teacher is already there giving their 100%. Let's give them a classroom that works. 🇮🇳 👉 Turn your intent into a verified physical asset today at TrueGiv.in #CSRIndia #Section80G #TaxSavingDonations #TransparentNGOs #ViksitBharat #SocialImpact #EducationInfrastructure #TrueGiv #DonateMoneyOnline
Experience & Education
Licenses & Certifications
Honors & Awards
-
Certificate of Merit
IIT-Kanpur
Awarded to the top 10% of the class, based on their GPA
-
Scholar Badge
DPS, R.K.Puram, New Delhi
For securing more than 80% marks in a year
Recommendations received
3 people have recommended Pushkar
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