The funnel was throwing away decisions users had already made

I owned the numbers, which meant I owned the arguments about them

I was the growth analytics owner for Flipkart's smartphone business. Six levers — EMI, exchange, buyback, GST, cancellations, SuperCoin — and one job underneath all of them: whatever the leadership review looked at each morning, I built and maintained.

That sounds like a reporting role. In practice it was a product one, because the reporting kept being wrong, and everything downstream of it was a decision someone was about to make.

The week it broke

I built the live growth dashboard for Big Billion Days 2023 — hourly and daily, all six levers, five people feeding planned units into it. It was the highest-traffic week of Flipkart's year.

It broke, publicly, for four days. Division errors across the unit-share column. Burn figures that didn't reconcile against the EMI dashboard. A timestamp that updated correctly while the data behind it didn't. By the first of September it was reporting EMI unit adoption at 82%, which is impossible, and the head of the business was asking me directly what was wrong with my tracker.

The cause was mundane: the raw sheets had grown past their allocated size, so formulas stopped copying into new rows. Silent, and only visible once the output was obviously absurd.

I fixed it that day. What mattered more was what I did after — rebuilding burn onto a separate tracker keyed on offer IDs so it would reconcile independently, and reporting the two things it still couldn't do rather than waiting to be asked.

That week is why I spent the next year on instrumentation.

Thirteen months arguing with a platform team

Euclid was Flipkart's central analytics platform. It had its own roadmap, its own priorities, and no obligation whatsoever to the smartphone business.

I started in January 2024 with a gap between their numbers and ours on the sales tracker, and established the visit-to-product-page tracking was broken on their side. Then that the marketing channels were mapped wrong for Shopsy. Then that the exchange funnel didn't exist in their schema at all and needed architectural work, not a config change.

None of that was mine to fix and none of it was theirs to prioritise. So the work was structuring it into something ownable — auditing which metrics existed in the schemas, getting engineering to check ingestion on the ones that didn't, chasing the remainder owner by owner, and escalating to central analytics when direct chasing stalled. Exchange funnel metrics landed in June. EMI tracking after that. The migration completed.

I have never had a better lesson in getting something built by a team that doesn't report to you and doesn't need you.

What the instrumentation showed

Once the exchange funnel was properly tracked, the drop-off was obvious.

Adding an exchange to an order took four steps on the product page: device brand and model, condition, policy, then a quoted price and discount. Users completed it. Their device details were saved.

But the selection flag was discarded the moment they navigated to another product page. So a customer who had already told us they wanted to exchange their phone — and given us every detail to do it — was asked to do the whole thing again on the next device they looked at.

Nobody had designed that. It was an artefact of how session state was handled, invisible until the funnel was instrumented well enough to show people dropping at a stage they had already cleared once.

The fix, and why it isn't a dark pattern

We persisted the opt-in for the rest of the browsing session. Deselecting stayed a single click, in the same place the offer had always appeared.

That distinction matters and I'd defend it in either direction. Defaulting someone into an exchange would be indefensible — it changes the price they see and commits them to handing a physical device to a courier. But the decision was already theirs. We weren't making it for them. We were stopping the product from forgetting it.

A separate arm tested an opt-in prompt between the product page and checkout for users already showing exchange intent.

Exchange adoption moved from roughly 12% to roughly 15% — about a 20% relative lift on a funnel carrying a large share of smartphone GMV.

The other half of the job was saying no

Twice that stands out.

In March 2024 exchange adoption dropped and I was asked for a root cause under a deadline. I found confirmation-to-order conversion down around 40%, ran city-level cuts, found some cities worse than others — and then reported that the city effect wasn't large enough to explain the decline. It wasn't the answer anyone wanted mid-RCA. The real cause turned out to be exchange prices falling, 11% in metros and 28% in smaller cities, which carry most of the volume.

In December 2024 I was handed buy-with-exchange numbers I couldn't reproduce. The order-status mix didn't match anything I could derive from the same tables. I ran the RCA, and the logic behind them was wrong.

What I took to Meesho

That the expensive failures aren't wrong decisions, they're decisions made confidently on numbers nobody checked. Flipkart taught me that by breaking in front of me during the biggest week of the year.

Six months after my last Big Billion Days, I was at Meesho building automated data-quality checks with alerting across the tables everyone else's decisions sat on, and arguing an engineering manager into paying to keep them running. I knew exactly what I was buying.