I'm Dheer, a product leader in Bengaluru. I've built B2C and B2B products from conception to scale: payments at Paytm, food ordering at Rebel Foods, gold loans at Rupeek and workforce software at BetterPlace. Right now I'm building hands-on with AI.
I try to work on interesting problems.Often quite hard.
I started on a factory floor, then spent eight years in Indian startups building for shopkeepers, people ordering food online, borrowers and blue collar workers. I'm drawn to problems where software has to meet the physical world: a QR code at a tea stall, a machine that turns jewellery into a loan in under ten minutes, a worker verified and onboarded before the first shift.
I'm on a deliberate sabbatical, building with AI every day, and open to the next hard problem, in whatever shape it comes.
Every role started with a problem worth losing sleep over. Here is each one, in my own words, and what came of it.
How do you make a factory produce more, and deliver on time, through better planning?
Three years in supply chain, SIOP and factory setup. Planning a production line is product thinking under hard constraints: lead times, capacity and people on shifts.
Then two years at SPJIMR, Mumbai, for a PGDM.
How do you make a roadside vendor creditworthy through the acceptance of digital payments?
Payment history was the first step to credit, so the work began with acceptance. Built a self-serve merchant onboarding platform, then drove activation with gamified, multilingual campaigns and real-time incentives, each targeted by auto-segmentation. Designed the tools and tracking for a national field sales force.
How do you offer convenient, affordable, quality meals for the many ways urban Indians eat?
Launched EatSure, which lets customers order from several of Rebel's brands in a single order. It consolidated individual brand websites, reduced reliance on aggregators and paved the way for Rebel Launcher, which incubates third-party restaurant brands. Shipped ML-driven recommendations and quick-cart checkout.
How do you offer the fastest, cheapest credit against gold jewellery to a vast India that lenders can't underwrite?
Pioneered the Gold Teller Machine, an industry-first hardware and software product for self-serve gold loans, with real-time KYC, AI-enabled fraud checks and automated secure jewellery storage. It processed over a thousand loans. Designed the lending stack behind bank and NBFC partnerships, and launched a hybrid secured and unsecured credit line.
How do you improve the employment experience and opportunities for blue collar workers?
Led seven PMs on a B2B HRMS platform for hiring, verifying, onboarding and managing blue and grey collar workers, directly or through vendors. Owned product, go-to-market and pricing. Brought fine-tuned LLMs and ML models into manual workflows, and led MyRocket, a job board for volume hiring that cut cost and time to hire by up to half. Also helped OkayGo, MyRobin and Troopers, the group's companies in India, Indonesia and Malaysia, shape their strategy and roadmaps.
What changes about building products when the model can reason, act and use tools?
A planned break to go deep on the modern AI stack: LLM architectures, agent harnesses, context engineering, orchestration and evaluation pipelines. I've also been mapping the vertical AI landscape, testing business models, feasibility and moats. And building.
Small, working products, each aimed at one specific problem. Writing the code myself, with AI coding agents, keeps my product judgment honest about what AI can and can't do yet.
Most SaaS onboarding fails at the moment of confusion, not for lack of documentation. Claro is an in-app guide that acts in that moment.
Ask by voice or text and an AI cursor walks you through nested menus step by step. It watches for signs of confusion, like long idle time or repeated clicks in one place, and offers help before you ask. Answers come from the product's own docs, and it asks a clarifying question when a request is ambiguous. Voice round trips stay under two seconds.
Many Indian professionals think in a mix of Hindi and English. Most dictation tools transcribe that literally. SenseIt turns intent into clean professional English.
A global hotkey records, and the refined text is typed straight into whatever field is active. It reads the context of the current app, so it can tag the right file in an editor or the right colleague in Slack.
An expense tracker I use every day, designed around the fewest possible keystrokes and no hosted data.
A parser categorises each message, and a model is called only when confidence is low, which keeps it inside free tiers. Entries are stored AES-256 encrypted in the user's own Google Drive app folder, with no database. A mobile-first dashboard shows trends and lets you edit any entry.
Better prompts come from knowing what to ask for. Forty-eight photography and film terms, from light to lens to movement, each with an interactive diagram.
A formula lab builds a prompt one decision at a time and flags contradictions, adapting the output for different image and video models. An A/B lab changes one variable at a time, and practice briefs are graded out of a hundred.
Formed from building, argued in public, and checked against the evidence below.
An agent is a model plus a harness: the prompts, tools, context management, verification and memory around it. Hold the model fixed and the harness alone can move a coding agent from outside the top 30 to the top 5. Harnesses are also tuned to one model's habits, which is why the labs that know a model first tend to ship the strongest harness for it. So the general-purpose layer belongs to the labs. What lasts is the domain layer: the tools, checks and permissions that encode what a model can't know.
So for builders: compete where a lab can't follow you, in a domain's tools, checks and workflows, not in a thin layer over someone else's harness.
Not every part of a harness is a moat. Much of what makes a wrapper work today is workaround: prompt chains, output-repair parsers, forced reasoning steps, hand-built task decomposition. Each model generation makes some of it unnecessary, but it rarely gets deleted. Nothing breaks when a workaround goes stale. It just keeps costing tokens and latency, needs re-testing at every model upgrade, and can stop a better model from doing what it now does on its own.
So for builders: treat workarounds as debt with an expiry date, audit what is still load-bearing at every model release, and spend what you save on the domain layer.
encodes an assumption about what the model can't do. The last row is a typical product.
The valuable layer is decision context: not just what happened, but why it was allowed, who approved it and which precedent applied. A popular thesis says vertical agents, sitting in the execution path, will own this "context graph". I doubt it. One agent sees one workflow, while real decisions draw on context from everywhere. Keeping that graph accurate takes governance an enterprise won't hand to a vendor. And context is already becoming portable between AI products.
So for builders: own proprietary data that is hard to synthesise, and be the best place to act on context the customer keeps.
I learn by building instruments for it. These are private projects, and I'm happy to demo them.
An interactive course in rhythm, harmony and style, from backbeats to Hindustani taal. Every sound, tabla and tanpura included, is synthesised live in the browser.
All 252 episodes of the Philosophize This! podcast, distilled into linked concepts and 250 thinkers, drawn as a map organised by the question each idea tries to answer.
A phone-sized chord reader and editor over the song list I keep in Google Sheets, because the Sheets app is unusable mid-practice. It writes straight back to the sheet.
Also a long-running self-study of theoretical physics and the mathematics underneath it, and an active investor in public markets.
Product judgment first, then the right tool for each step, from a spec to a shipped prototype.
Claude Code
Codex
Cursor
LangGraph
Hugging Face
Python and SQLPGDM, S. P. Jain Institute of Management and Research (SPJIMR)Mumbai, 2014 to 2016
B.Tech, National Institute of Technology (NIT Trichy)Tiruchirappalli, 2007 to 2011
Strategy, discovery, go-to-market and pricing across B2C and B2B. Platform and marketplace design. Machine learning products, RAG, agent orchestration and LLM evals.
I'm open to the right next thing: a full-time leadership role, a consulting engagement, or advising a team that's building with AI.