Redesigning Pathao's platform home for discovery, personalization, and visual clarity.
The home was a static grid that treated every user the same, so most of the ecosystem stayed invisible. I reframed the problem, rebuilt the home around relevance, and built a Stable Diffusion + GPT-4 pipeline to produce the icons that made it possible.
Pathao's home had to serve a whole ecosystem, and served none of it well. Every session opened the same grid, in the same order, for everyone — so the app kept teaching people that it was a taxi app with extra tabs.
The same broken home failed three groups in three different ways.
New / exploring users couldn't self-guide. Anything filed under "More" was effectively invisible.
Frequent users got no adaptation. The same irrelevant grid every session; shortcuts like "Ride to Home" or "Order Again" arrived without context.
Cost-conscious users missed the offers meant for them. Points sat at ~0.5% engagement — people trusted in-service deals over a buried tab.
I framed the problem from user interviews, stakeholder input, and platform event analytics rather than instinct.
Discovery happened opportunistically — after a ride, mid food order — never deliberately. People hunted for services under "More" and gave up. Pay Later sat next to Pathao Pay and was consistently misread as the same thing. Asked what they wanted, they described a home that changed with them.
Rides and food dominated every session. Other sections were skipped outright. Shortcuts were noticed but not trusted enough to tap from home.
Low engagement despite prominent placement
Users leaving home to find what should already be there
Buried inside profile
Nearly invisible
The problem frame set the bar. These kept the solution honest to it.
Personalization earns its place only when it surfaces what someone actually needs. Fewer, better tiles — never more of them.
A service buried under "More" doesn't exist. Hierarchy and clarity beat visual novelty every time.
Generated output is a candidate, never a commit. The design system and human review are the gate.
Every produced icon still has to read as unmistakably Pathao — on-grid, on-brand, and local rather than stock.
A personal home that guesses wrong costs more than a static one. Explainable, correctable, never uncanny.
Research synthesis
Interviews, stakeholders, analytics
Problem frame
Core statement, themes, constraints
Design goal
Personalized, discoverable home
Home IA
Slots, hierarchy, navigation
Icon pipeline
Stable Diffusion + GPT-4
Guardrails
Prompts, limits, human review
Validation
Usability + generation quality bar
Iteration
Tune against real engagement
Production
Rollout, product by product
Intent — read behavior, time, location, and history into what someone is likely here to do right now.
Ranking & composition — choose which services, merchants, and offers appear, and in what order, for this specific home.
Language — frame each tile with copy that stays contextual and concise instead of generic marketing.
Personalization was a product and IA problem, not a model problem: turning signals into a home that feels deliberate for each person.
Moving between services stopped being a context reset and became a continuation of one journey — the same hierarchy, the same gestures, the same sense of place, whichever service the home led with.
Personalization without icons is empty frames. Icons without a system are a design queue that never ends. AI earned its place on the second problem only — the long tail of on-brand icons no team could draw one at a time.
Thousands of surfaces. Merchants, categories, and offers all needed tiles. Photography and hand illustration were never going to cover them.
One visual language. Every generated icon still had to read as Pathao — on-grid, on-brand, local, not a stock collage.
Ship without pausing. Squads kept shipping. The pipeline had to slot into the existing home rather than demand a rebuild.
Design stays the gate. Models propose. Prompt rules, review, and the principles decide what reaches a user.
The pipeline paired Stable Diffusion for output with GPT-4 for prompting and iteration — and kept design as the thing that decided what shipped.
Prompts, style references, and post-generation rules kept every icon inside one visual language.
I started from the set users already knew, so the new work would feel like a continuation rather than a reset.
During my iteration cycle, I refined over 1500 icons, ensuring they all followed a cohesive visual language, guided by prompts, style references, and post-generation guidelines.
Once the visual direction held up across enough subjects, the pipeline could keep feeding the personalized home — same gate, more surfaces, no one-off fixes.
Getting there meant refining over 1,500 icons for a cohesive look, guided by prompts and style references.
Context-aware generative image models changed how I build a scalable visual framework — the bottleneck moved from production to judgement.
They also made cross-functional work faster: teams could see options instead of waiting on a queue, and review moved earlier.
The result is one family that scales: consistent construction, consistent weight, and enough range to cover services that didn't exist yet.
Every icon shipped through the same gate — on-grid, on-brand, and reviewed by a person before it ever reached a user.
Personalization is easy to claim and hard to prove, so the work was held to engagement data across the whole ecosystem.
| Product | Monthly events | Users | What it showed |
|---|---|---|---|
| Car | 2.5M | 605K | Mature anchor — deeper engagement per user |
| Food | 1.8M | 262K | Grew as discovery routed ride users in |
| Parcel | 652K | 253K | More completed per session |
| Courier | 154K | 24K | Smaller audience, markedly deeper reliance |
| Tong | 61K | 34K | Fast adoption once surfaced on home |
The signal wasn't more taps on home. It was discovery compounding into depth — as the home surfaced secondary services, engagement per user rose almost everywhere downstream.
The ecosystem's anchor, and the hardest test — mature surfaces rarely move.
A mature product used more deeply — engagement per user climbed on the surface with the least room to grow.
The clearest proof that surfacing a service accelerates it.
Once the personalized home put Tong in front of the right people, a small vertical grew on every axis at once.
Discovery routed ride users into Food and they stayed — growth without regression.
Shipment visibility moved from an occasional check to a routine part of the journey.
People completed more per session — a clean signal of stronger feature adoption once found.
The instinct with AI is to point it at the interface. The leverage was pointing it at the one thing design couldn't scale by hand — icon production for the long tail — while personalization and discoverability stayed product decisions, with design holding the gate on everything that reached a user.
AI didn't replace design decisions. It made more of them possible — and made the guardrails the most important thing I designed.
More projects are being documented.
Other case studies are in progress. Check back as they ship — or head home for featured work and contact.
