// Byju's Wiz AI
Lead UI/UX Designer on Wiz AI at Byju's. An AI doubt-solver for ~5M monthly active learners. Shipped with PM Anukrit Jain and AVP Himanshu Nimje.
## Problem
Students got stuck. The existing human-led doubt flow was slow, transactional, and expensive per session. ~80% of students who wanted instant resolution said the experience wasn't good enough.
A doubt is not a curiosity. It's friction. The fix had to be fast, cheap, and trustworthy.
// Problem overview
## Approach — the 70/30 rule
Hard product rule: 70% of every answer came from Byju's existing content library, 30% was AI-generated to personalize and fill gaps.
Library content was cheaper and already vetted. The rule cut daily cost by ~71%. That constraint didn't just live in the backend — the UI had to enforce it without feeling restrictive.
// Wiz AI flow
## Decision 1 — Library vs AI
Early designs surfaced "Generate with AI" as the dominant action. It felt intuitive. It also pushed users toward the most expensive path and framed library content as a fallback.
// Rejected direction
Students, especially during exam prep, trusted structured explanations more than "generated" ones. The word AI made some cautious.
Final pattern: library matches surface first. "Generate with AI" is visible, equal weight, not dominant. Both live on the same surface.
## Decision 2 — Latency is visible
Every answer ran through multiple steps. ~6–8% were revised after review. Lower latency meant higher infra cost.
Instead of hiding the tension, we showed it. Shimmer, streaming text, sticky loader that persists on scroll. When a reviewed answer replaced an earlier version, a clear refresh prompt — not a silent swap.
// Generating state — shimmer, streaming text, visible progress. // Sticky loader — stays visible as the student scrolls.## Decision 3 — Scoped follow-ups
Most AI chats keep the whole thread as context. Smart. Expensive at scale.
We made students pick the answer they wanted to follow up on. Only that answer is used as context. Conversations stayed focused. Tokens stayed bounded. Repeated doubts went down.
## Impact
71%
cost reduction
5M+
monthly active learners
80%
satisfaction score
85%
faster resolution
Before Wiz AI:
→ ~1,014 daily doubt sessions went through human-led DoC → ₹50 per session → ~₹50,700 total daily cost
After Wiz AI:
→ 207 sessions still needed human DoC → 808 handled by Wiz AI at ~₹5.5/session → ~₹14,800 total daily cost — a 70.9% reduction
Satisfaction: 79.6% of Wiz AI users rated it "Good" vs 69.7% for traditional DoC. ~14% relative improvement.
## Close
Library-first, explicit follow-ups, visible latency. Three design decisions, one cost curve. The product felt calm because it was calm.