I build B2B SaaS products that
drive better business decisions.
Five years of product across enterprise data platforms, industrial SaaS and public-sector operations. My job is usually the same one. Somebody rebuilds a spreadsheet every Monday because they do not trust the system, and I have to build the thing they trust instead. I start with discovery, instrument it on day one, and close every release out with a number.
Blank page to five signed enterprise accounts as the only PM. I agreed the data model and metric dictionary with finance before anyone designed a screen, then shipped seven modules against it.
Three shipped AI surfaces and an agent prototype. In all of them the model writes the sentence and the pipeline produces the number, so a wrong figure is not something a prompt has to prevent. Thresholds set for precision. A human left in the loop, because somebody has to own a wrong answer.
Instrument first, then argue. Every number on this page says how it was measured: the instrument, the window, what it was compared against. Ranking friction by users lost ended a two-quarter debate and moved conversion, adoption and revenue.
I settle what a number means before anyone designs a screen.
Five years across three very different B2B products. An enterprise data intelligence platform, an industrial operating system for manufacturing plants, and a public data programme run at state scale. Different markets, same job. Find where the decision actually gets made, work out why people trust their own spreadsheet more than the system, and build the thing that replaces it.
Most recently I took a platform from a blank page to five signed enterprise accounts as its only PM. I agreed the data model and metric dictionary with finance before anyone designed a screen, then shipped seven modules against it. Before that I owned a multi-module roadmap through 100+ features, which lifted adoption 35% year on year and annual revenue 20%.
The AI work is what I care most about now. I treat it as a production judgement call, not a feature label. Four of the eight cases here are AI surfaces. In one, generation only runs over pre-computed indices, so the read-out cannot be confidently wrong. In another, the alert threshold is set for precision, because a channel people mute is a dead feature. In the third, the agent has exactly one typed tool. It turns a question into a filter state and never calculates anything itself. Which model sits behind it interests me far less than where the guardrail sits.
I write these case studies the way I would defend a decision in a review. What the constraint was, which option I turned down, what the tradeoff cost me, the number it moved, and what I would do differently.
Himanshu RanaProduct ManagerGurugram, IndiaDiscovery with the people doing the manual version. The spreadsheet they built for themselves is the specification — including the parts they are embarrassed by.
Metric dictionary and entity grain settled with finance and data before a screen is designed. Numbers that disagree across screens kill adoption faster than missing features.
Each release completes a whole decision rather than a third of three. Instrumented on day one, measured against the manual baseline it replaced.
Eight cases, eight decisions.
Four from the platform I own now, including the agent prototype. Three from an industrial SaaS roadmap. One from operations at state scale. Open any card for the full write-up: the constraints, the option I rejected, what the tradeoff cost, the measured impact, and what I would do differently. Product screens come from the live tool with commercial figures blurred.
On verifying these numbers. Every case carries a How this was measured note saying what the instrument was, what window it covers and what it was compared against. Judge the method, not just the figure. You will see percentages instead of absolute revenue and volumes, because those are commercially confidential. Where a number was reported to me instead of computed by me, the note says so. Happy to walk through the underlying boards, dashboards and PRDs in an interview.
- Context — the commercial situation
- The problem — what was actually broken, not the symptom
- Constraints — what I could not change
- Approach & key decisions — including the option I rejected and what it cost me
- How it shipped
- Impact — the number it moved
- How it was measured — instrument, window, comparison basis
- What I'd do differently
Where it was built.
Three organisations at three very different scales. An enterprise platform built from nothing, a multi-module industrial roadmap, and a public programme running across 82 towns.
Certification
CertificationWhat I work with.
◈Product
▤Data
✦AI & Gen AI
◎Domains shipped in
Let's talk.
Open to product roles and to interesting conversations. 0→1 builds, data platforms, AI inside a workflow, and anything where the current answer is a spreadsheet somebody rebuilds every Monday.