I’ve spent six years deleting the interface

When we sat with users in small towns, some of the friction was humbling. The cart icon meant nothing to people who had never pushed a trolley through a supermarket. Typing into a search bar was slow and felt unnatural.

E-commerce was designed by people who already knew how to shop online. For the next few hundred million users, the familiar model is the neighbourhood shop: you describe what you need, the shopkeeper asks a question or two and shows you three things. We weren’t trying to teach people the internet’s metaphors. We were trying to bring the shop’s conversation online.

Many rural users couldn’t fill our address form. They knew where they lived; that just isn’t how an address exists in their heads. They know the village, a landmark and whose house it is.

Voice let us ask the way a person would. The hard part was knowing what to ask next and when to stop, so we built that as an explicit state machine tied to geocoding. When users fail a form, check whether the form matches how they hold the information before you redesign it.

Voice is intrusive. It takes over someone’s attention in a way a screen doesn’t, so a voice turn has to give more than the screen would have.

Every turn also costs money. In a business where an order is worth a few hundred rupees, a chatty assistant can lose money on the very users it was built to win. Fewer, better turns was the right rule for experience and for unit economics at once.

If adoption depends on users discovering a new way to use the app by themselves, it won’t happen. Vaani reached 39% adoption because it introduced itself during onboarding, and offered help when behaviour showed someone was stuck: a stalled payment, a search that came back empty. The assistant should show up when it’s useful.

Before Vaani we simplified screens, added explainer videos and tried audio guidance. Each moved activation a little and then flattened out for rural cohorts. Small wins that stop compounding were the signal that we were executing an idea with a ceiling. That’s when an expensive change of interface becomes worth paying for.

Cost per conversation decided whether we could reach every user or only a pilot cohort. Vaani’s voice stack ended up 22× cheaper than the default cloud setup, mostly from moving speech on-device, a choice that came out of benchmarking rather than conviction. A PM who treats inference cost as engineering’s problem ships something that works in the demo and gets rolled back at the budget review.

Real speech is noisy: mixed languages, background sound, cheap microphones. A pipeline that assumes a clean transcript fails exactly where it’s needed most. We built a salvage layer that recovers intent from noisy transcription and keeps what the assistant says on a controlled script. A wrong word on screen is a typo. A wrong word spoken aloud with confidence breaks trust.

You can’t always make the model faster, but you can make the wait feel different. We let the screen respond and move while the voice catches up, instead of leaving people in silence. For the phone agent, latency mattered more than benchmark quality when choosing the model and inference partner.

Most evals ask whether the model answered well. That can’t tell you whether the conversation led to an order. We joined clickstream to conversation logs so every conversation could be traced to what the user did next. The rest is unglamorous: golden datasets, sampling heuristics, people reading transcripts every week.

The address flow could have been one big prompt. It worked better as a state machine with the model doing what it’s good at, which is understanding a messy answer. Use the model for language and keep the logic where you can see it, especially where a mistake sends a parcel to the wrong village.

Meesho’s Agentic Marketplace Platform exists because Vaani came first and taught us what memory, evaluation and orchestration needed to do. A platform designed before a real product has used it tends to be general in the wrong ways. Two rules we hold to: stay model-agnostic, and make build-vs-buy calls layer by layer.

In late 2022 I scored every GenAI idea on three things: the size of the lever, closeness to the core business, and how well today’s technology fits the problem. The third question does most of the work. Plenty of valuable problems are a poor fit for what models can do right now.

The move from chat to voice came from a hackathon prototype that leadership could phone and hear. With AI products, how the thing feels is much of the argument, and you can’t put that in slides. To change a big decision, build the smallest thing someone can try.

After pitching our chat agent to 30–40 startups I thought selling it wouldn’t work. Small businesses lacked the volume, incumbents owned the integrations, and chat was too easy to copy. Voice over telephony avoided all three because it was harder: latency, Indian languages, telephony plumbing. The difficulty was the moat.

Small and long-lived, because know-how about AI products builds up in people. Someone who reads transcripts every week, because dashboards tell you that something changed and transcripts tell you why. Field time for everyone: engineers who have watched a user struggle make different trade-offs. And permission to kill things cheaply, without it being a career risk.

In AI products the prompt, the eval set and the conversation design are the spec, in a form that runs. A PM who hands the prompts to engineering has handed off the product. I write prompts, read failure cases and own the golden datasets, because that’s where decisions about what the product does are actually made.