Sharanya Ramesh

ChatGPT is not just for casual chats; it can serve as your go-to resource for Product Analytics!. Metrics don’t fail you. Vague thinking about metrics does. This DYK chapter is about moving from “we shipped it” to “it worked (or didn’t) — and here’s why.” Defining success properly. Separating leading from lagging indicators. Killing vanity metrics. Interpreting changes without panic. Used intentionally, ChatGPT becomes your Product Analytics Lead — helping you measure what actually moves the business. Save these prompts. Better metrics = better decisions. #ProductAnalytics #ProductManagement #GrowthThinking #AIPoweredPM #DYKSeries #PMCraft #DataImmersionHub

7

#Tokenization just quietly became the new baseline for card payments – and Visa’s latest earnings call #FY26Q1 makes that crystal clear. A few signals that stand out: - Network tokens have #scaled to massive volume (17.5B) now far beyond the underlying physical cards. - Tokenized transactions are consistently delivering better #outcomes – higher approval rates and lower fraud. - The biggest digital merchants are already operating as if tokens are the default credential, not an optional add‑on. - New themes like #AI‑driven/#agentic commerce are being architected on top of tokens, treating them as the payment primitive. The shift is subtle but important: Tokenization is no longer: “Nice security layer we’ll integrate later.” Tokenization is now: “The minimum viable substrate for any serious digital payments roadmap.” https://lnkd.in/ghWZWAvW

64

1 Comment

Agentic Payments: What changes for me as a Product Manager? UPI is moving from “Tap to Pay” to “Set the rules, let AI pay.” GFF (Global Fintech Fest) 2026 has made one thing pretty clear: Agentic AI is moving from a buzzword to actual fintech products. Today: User → selects → approves → pays Tomorrow: User → sets rules → AI decides → UPI executes NPCI’s Unified Agent Protocol (UAP) is being designed around this exact shift, allowing agents to make payments within user-defined permissions and limits. So I started thinking: What actually changes for me as a Product Manager? The problem statement is no longer just “How do we make payments faster?” It becomes: “Which decisions are users comfortable giving up?” I’d start with: → Grocery reorders → Recurring bills → “Buy when price < ₹X” → Predictable, low-risk purchases Because frequency + predictability + low consideration = easier trust. And then comes the real product problem: What happens when the AI gets it wrong? Who sets the limit? Can I revoke it instantly? Can I restrict certain merchants? Can I see why the agent made the purchase? Suddenly, trust isn’t a compliance layer. Trust is the product. And we’re already seeing the pieces come together. PhonePe is building intent-led experiences, while Razorpay is pushing agentic payments, Agent Studio and its payments-focused AI model, Vulcan. So here’s the question I’d ask as a PM: If the agent owns the decision, what happens to the payment app’s role? Does UPI become the execution layer? Does the AI agent become the decision layer? And eventually… Are the best fintech products the ones that help users make fewer decisions, rather than complete more transactions? That’s the shift I’m watching. Genuinely looking forward to the Global Fintech Fest 2026 📈✨ #ProductManagement #Fintech #UPI #AgenticAI #Payments #ProductStrategy #GFF2026

9

With the advent of AI tech challenged PMs will soon become history and esp. important is to understand constructs like **Tokenization** which are important for all language models be it the classical NLP models or the LLM models. It also helps understand why **byte pair encoding** works for tokenization for pre-training of LLMs. The idea behind this is simple, for any algorithms to process natural language which is continuous and unstructured, you need to divvy up this text into more atomic units. If you don't split this into more atomic units, how will you convert it into structured inputs to run any statistical analysis/more advanced ML/AI. Eg. take the simple case of sentiment analysis through bag of words. Take reviews of a movie "I love this movie" → Positive (1) "I hate this movie" → Negative (0) "This movie is great" → Positive (1) "This movie is bad" → Negative (0) Unless you split this corpus of text into say individual words and attach some numerical representation to the individual words how will you represent the sentiment analysis problem as a mathematical problem which then machines can solve. This entire process is called tokenization. So in this case you can build the master vocabulary as no of unique words in the text corpus. So you divvy up the words and find unique words. Vocabulary = ["I", "love", "hate", "this", "movie", "is", "great", "bad"] Then you represent each sentence as a simple vector in which the words are features. (1s and 0s features represent whether word is present or not) Vocab I love hate this movie is great bad Review I love this movie(1) 1 1 0 1 1 0 0 0 I hate this movie(0) 1 0 1 1 1 0 0 0 Now it becomes a classical regression (logistic regression) problem where sentiment is 'output variable' and the unique words represented as 0/1 (feature vector) become input variables and you can fit a curve. Thus it becomes a mathematical problem similar to classic regression where let's say you are trying to predict house prices basis pin code, house size, type of house, amenities etc. Now the fun bit is the same tokenization construct is used when doing the pre-training for LLMs. ***The only difference is the tokenization happens at a subword level and uses an algorithm called byte pair encoding.***** Why subword because to be able to understand and practically speak like humans, LLMs need to understand that "boy" and "boys" are not necessarily different words or "tokenization" and "memorization" uses the same root suffix "ization" Idea remains the same you split the words at subword level and attach mathematical representations and then convert these tokens to embeddings for LLM pre-training. Follow me https://lnkd.in/gHweK_q7 AND https://lnkd.in/ggUKftFq For tech mastery - https://lnkd.in/gR2fnkAa

17

2 Comments

Evaluation and Benchmarks - How to Measure Agent Performance You built an agent. How do you know it actually works? Measuring agent performance is fundamentally harder than measuring model performance. Agents are multi-step systems that use tools, interact with environments, and produce real-world effects. A wrong step early cascades through everything that follows. Traditional accuracy metrics do not capture this complexity. The current benchmark landscape: - SWE-bench: resolve real GitHub issues (code agents) - WebArena: complete browser-based tasks (web agents) - GAIA: multi-step reasoning across tools and information sources - OSWorld: operating system-level task completion Four metrics that matter in practice: - Task success rate: binary completion (did it finish the job?) - Step efficiency: how many steps vs the optimal path - Tool accuracy: right tool, right parameters, right timing - Cost per task: tokens + API calls + compute time The evaluation gap most teams miss: benchmark performance does not equal production performance. An agent scoring 80% on SWE-bench may fail on your specific codebase patterns that benchmarks never tested. The solution is layered evaluation: - Component evals find individual bugs (tool selection, planning, memory) - End-to-end evals find integration failures - Offline evals enable fast iteration during development - Online evals monitor real production performance Building custom evals: define 50-100 representative tasks, create scoring rubrics, use automated graders (a second LLM evaluating the first), and track metrics over every deployment. How are you evaluating your agents beyond demos? Follow for daily insights - where blockchain meets AI, one satisfying swipe at a time. #AgenticAI #AIEvaluation #Benchmarks #SWEbench #AIEngineering

1

Amazon Pay India has launched fixed deposit (FD) investing directly inside its app, enabling users to open FDs starting from just ₹1,000 — without needing a separate savings account with the partner institution. This service is powered through partnerships with two NBFCs (Shriram Finance and Bajaj Finance) and five banks including Shivalik Small Finance Bank, Suryoday Small Finance Bank, South Indian Bank, Slice and Utkarsh Small Finance Bank. 👉 Key highlights: 📈 Up to ~8% interest p.a. on FDs — competitive for safe, fixed-income savings. 👴 Additional benefits for senior citizens and women investors via enhanced rates. 🛡️ DICGC insurance coverage up to ₹5 lakh per depositor per bank — bringing the same safety assurance as traditional bank FDs. 📱 Fully digital process inside the Amazon Pay app — compare options, choose tenure, and complete investment seamlessly. This move marks a meaningful evolution of Amazon Pay from a payments tool to a broader financial services platform, blending convenience with wealth creation. 📌 Why this is significant ⚡ It lowers barriers to fixed-income investing for millions who already use Amazon Pay for UPI, bill payments, shopping and credit solutions. ⚡ It brings bank-grade financial products into the hands of everyday users without the friction of traditional banking interfaces. ⚡ It signals intensifying competition in India’s fintech landscape — where payment apps increasingly blend payments, credit, savings, and wealth within a single ecosystem. As the digital finance market matures, the blending of transactional and investment rails will redefine how users save, spend and grow their money — right from the apps they already use. #AmazonPay #Fintech #DigitalFinance #FixedDeposits #Investment #Savings #WealthTech #DigitalIndia #CustomerExperience #FinancialInclusion #Innovation https://lnkd.in/dr3xDKkS

3