Prerna
Kapoor.

At-home eye care is especially useful for customers who would rather call than navigate an app. Senior citizens and other less tech-confident customers may want a home eye test or frame trial, but need a low-friction way to complete the booking independently.

Charlie creates that access path: a customer can make a simple phone-style call, speak naturally, and have the agent book a visit on their behalf. The product challenge was to make this feel reassuringly simple while keeping the booking process precise and reliable.

Charlie uses voice only as the interface. The deterministic state machine decides what happens next, so every step is explicit, validated, and traceable.

A deterministic state machine owns every booking decision. An optional model only helps interpret the caller's current answer.

Charlie is a reusable assisted-action pattern for apps where a customer needs to complete a real-world task but would benefit from a phone-first, guided route.

Four in five Home Trial landing-page visitors left at step one, before selecting an address or slot. We called people who had shown intent but abandoned there to understand the barrier.

With no say over which frames showed up, booking felt like a gamble. A store visit felt certain. This wasn't a distance problem or a trust-in-service problem; it was a control problem.

The hypothesis: it wasn't more frames customers needed, it was the act of choosing. So we tested that in isolation. Using an existing Lenskart product listing page, we let customers browse and "select" frames for their trial; from the same ~150 frames the agent was already carrying. Nothing changed in the backend. The agent showed up with the exact same kit as before, just now hand-picked by the customer on paper.

Launching home trials was the easy half. Scaling them meant going beyond a dedicated agent workforce; every Lenskart store needed to become a home trial hub.

Making that work meant answering a harder question first: which store owns which customer?

Wanna see how we solved that?

When a Lenskart customer buys in-store, a staff member is right there to nudge them toward a repair or exchange if something's slightly off; fit, alignment, comfort. Online customers get none of that. The order lands at their door, something feels wrong, and with no one to talk them through it, they go straight to a return.

The fix wasn't just closing that gap. It was making every store accountable for a defined territory; every customer in it, online, offline, or home trial, becomes that store's responsibility.

All three sound reasonable. All three try to infer which store a customer has a relationship with. The data said that's the wrong question entirely.

Polygon attribution replaced a blunt radius rule with a capacity-aware service territory. Inspired by Walmart's hex-grid catchment model and built from 0.1km² tiles, it assigns every delivery address to one accountable store — no overlap, no dead zones, and no inference about customer loyalty. Territories change only when the store network changes, making ownership auditable and targets explainable.

We built a unified "Your Territory" view inside the store POS; every store sees its polygon, its online and walk-in orders together, and one set of targets across both. Ownership meant giving stores real tools to act on it:

Health Files let customers upload their lab reports to the platform. Almost nobody did. Rather than optimising the upload flow, we went back to customers to understand why.

Lab reports from different labs use different names, units, and formats for the same parameter. We had to standardise before we could explain anything.

Every parameter explained in plain English; what it means, where you stand, what's critical, how to improve it; plus a historical trend compiled across all past reports.

Doctor listings had no real ranking logic; sort order was effectively random. Doctors with poor attendance records ranked identically to doctors who consistently showed up on time, and doctors with incomplete profiles — missing degree, years of experience, description, or even a photo — could still surface in the first few scrolls, ahead of doctors with complete, trustworthy profiles.

Before rebuilding the ranking, we identified the signals patients used to choose a doctor and treated the weighting as a product hypothesis to validate through live behaviour — not as a permanent truth baked into the system.

Rather than averaging every signal into one flat score, we split the model into two layers; preference signals and trust signals aren't the same kind of thing. A highly experienced doctor with a poor attendance record shouldn't be able to average their way to a high ranking — reliability has to gate visibility, not just adjust it.

Attendance % and Profile Quality combine into a single multiplier. A doctor with poor attendance or an incomplete, unprofessional profile has their entire score pulled down proportionally — no amount of experience or strong reviews can fully compensate for unreliability.

A complete profile isn't necessarily a trustworthy one; a doctor could fill every field with a blurry, badly cropped, unsmiling photo. So beyond checking whether fields were filled in, we bulk-cleaned photo backgrounds across the entire doctor directory for visual consistency, and scored photo professionalism directly on expression, blur, and framing (full face captured vs. zoomed-out or cropped).