Work

What I built, and what it cost.

The number first. Then the situation, what was hard, what got built, and what I would do differently.

1,300stores, more than 50 countries

XY Retail · 2026 —

AI on a fashion platform that already has stores

1,300stores, more than 50 countries

The situation

XY Retail is a unified commerce platform for global fashion brands. The stores, merchandising, demand and supply chain already exist. What they did not have was AI that could sit on that stack without becoming a demo nobody used on the shop floor.

What was actually hard

Fashion retail is merchandising, demand forecasting, supply chain and personalisation — work that already has owners. An agent that looks clever in headquarters and fails in a store is the whole problem. The platform is live in more than fifty countries; there is no maintenance window for a clever rewrite.

What I built

AI-driven platform capability across the retail continuum the stores already run — demand, merchandising, supply chain and personalisation — with the same gate I use everywhere else: a write waits for a person.

What it produced

This is the current engagement, not a closed case. The test is whether a store can use it on a Tuesday, across 1,300 locations in more than 50 countries — not whether the demo landed.

What I’d do differentlyI am watching the same miss I made on the lending line: treating the platform half as the work, when the stores are the constraint. Operations and merchandising stay in the room from week one this time.

Moneycontrol / Network18 · 2024–2025

A lending business inside a media company

US$23Mmonthly disbursals, from a standing start

The situation

Moneycontrol had a large, financially engaged audience and no way to serve them beyond content. The group wanted a revenue line that used what it already knew.

What was actually hard

Lending is a regulated business with credit risk, partner integrations and reconciliation obligations that a media organisation has no muscle for. The technology was the smaller half of the problem.

What I built

The platform, the partner integrations and the operational tooling from a standing start — origination, disbursal, servicing and reconciliation.

What it produced

Roughly ₹200 crore (about US$23 million) in loan disbursals per month.

What I’d do differentlyI would have put risk and operations in the room in week one, not after the platform looked ready. The integrations were always going to be the work. We spent too long making the obvious half look finished.

Moneycontrol · 2024–2025

Moving a national news platform to Google Cloud

20M → 80Mmonthly users, no maintenance window

The situation

Moneycontrol ran on on-premise infrastructure that could not absorb the traffic growth the business was planning for.

What was actually hard

Migrating without a maintenance window, on a platform where a market-moving news event is exactly when you cannot be down.

What I built

A full migration to Google Cloud with the platform re-architected for elastic scale and resilience — caching and in-memory layers, streaming pipelines, and a data warehouse to replace batch reporting.

What it produced

Monthly traffic grew from 20 million to 80 million as the publication expanded from business news into a multi-domain platform.

What I’d do differentlyI would have moved the cache and the warehouse first. Those were what made a no-window migration survivable. The application servers were the visible work, and they were not the constraint.

Wellness Forever · 2022–2024

Predicting disease from twenty years of pharmacy data

75%outbreak accuracy, 400+ stores

The situation

India’s third-largest pharmacy chain held two decades of prescription data across 400+ stores and used it for almost nothing.

What was actually hard

Prescription data is noisy, seasonal and privacy-sensitive. The useful signal is at postcode level, which is precisely where the sample sizes get thin.

What I built

Custom models producing a hyperlocal disease profile by postcode, plus an epidemic-prediction platform reading medication purchase patterns to forecast local outbreaks.

What it produced

75% accuracy on outbreak prediction, applied to demand forecasting, store site selection and supply-chain shortage prediction. Contributed 3–4% revenue growth. Recognised with CTO of the Year (HealthTech), January 2024.

What I’d do differentlyThe model was confident at postcode grain before the sample could support it. I would have published the uncertainty with the forecast from the first week, not after the first miss taught us to.

Ask me about any of these LinkedIn