Designing a system where disagreement becomes training data. Not noise.
WordSmart Annotator Workspace
The question wasn't "how do we annotate Indic content at scale?"
It was "why does every existing approach produce culturally flat data?" Answering that question — and designing around the real answer — is what this project was.
MY ROLE
Lead UX/UI Designer (research, IA, interaction, visual system, prototyping)
COMPANY
Wordwise Language Labs
TEAM
3 designer, 4 engineers, 1 PM,
1 AI-ML engineer.
TURNAROUND TIME
16 weeks, 8 two-week sprints · Figma, FigJam, Maze, Dovetail, Notion, React + Tailwind
MY ROLE
Lead UX/UI Designer (research, IA, interaction, visual system, prototyping)
TEAM
3 designer, 4 engineers, 1 PM,
1 AI-ML engineer.
COMPANY
Wordwise Language Labs
TURNAROUND TIME
16 weeks, 8 two-week sprints · Figma, FigJam, Maze, Dovetail, Notion, React + Tailwind
MY ROLE
Lead UX/UI Designer (research, IA, interaction, visual system, prototyping)
TEAM
3 designer, 4 engineers, 1 PM,
1 AI-ML engineer.
COMPANY
Wordwise Language Labs
TURNAROUND TIME
16 weeks, 8 two-week sprints · Figma, FigJam, Maze, Dovetail, Notion, React + Tailwind
What is WordsSmart?
What is WordsSmart?
What is WordsSmart?
WordSmart is a cultural annotation platform that produces the training data India's sovereign AI models actually need — not grammatically correct Indic text, but regionally authentic content verified by matched cultural experts, with a built-in audit trail for government compliance.
WordSmart is a cultural annotation platform that produces the training data India's sovereign AI models actually need — not grammatically correct Indic text, but regionally authentic content verified by matched cultural experts, with a built-in audit trail for government compliance.
WordSmart is a cultural annotation platform that produces the training data India's sovereign AI models actually need — not grammatically correct Indic text, but regionally authentic content verified by matched cultural experts, with a built-in audit trail for government compliance.
Impact
Impact
Impact
60%
60%
Reduce Switching
Reduce Switching
Apple Geniuses and Technicians no longer need to navigate between multiple articles during repairs
Apple Geniuses and Technicians no longer need to navigate between multiple articles during repairs
85%
85%
Chat Success
Chat Success
Users find the correct information every time they interact with the AI Assistant
Users find the correct information every time they interact with the AI Assistant
90%
90%
SCAN TIME
SCAN TIME
Technicians find related repair articles and documents significantly quicker.
Technicians find related repair articles and documents significantly quicker.
Discovery
Discovery
Discovery
Semi-Structured Interviews of 40+ Technicians Globally
Semi-Structured Interviews of 40+ Technicians Globally
We conducted 40+ semi-structured interviews with technicians across global markets to understand their roles, repair workflows, relationship with AI, trust expectations, and preferred interaction patterns. The research also included usability observation of technicians navigating GSX during real repair scenarios.
We conducted 40+ semi-structured interviews with technicians across global markets to understand their roles, repair workflows, relationship with AI, trust expectations, and preferred interaction patterns. The research also included usability observation of technicians navigating GSX during real repair scenarios.
The synthesis revealed recurring friction points across the workflow:
The synthesis revealed recurring friction points across the workflow:
Fragmented information: Technicians often stitched together 3–5 articles for a single repair, working across 3–7 parallel tabs.
Search-first dependency: Article titles frequently failed to reflect their actual content, forcing technicians to repeatedly search for information.
Low video engagement: Repair videos were often abandoned when lengthy introductions delayed access to the required step beyond approximately 30 seconds.
Disconnect from real-world repairs: Training footage often featured pristine devices, which did not adequately reflect the worn, damaged, or previously repaired devices technicians encountered in practice.
Fragmented information: Technicians often stitched together 3–5 articles for a single repair, working across 3–7 parallel tabs.
Search-first dependency: Article titles frequently failed to reflect their actual content, forcing technicians to repeatedly search for information.
Low video engagement: Repair videos were often abandoned when lengthy introductions delayed access to the required step beyond approximately 30 seconds.
Disconnect from real-world repairs: Training footage often featured pristine devices, which did not adequately reflect the worn, damaged, or previously repaired devices technicians encountered in practice.
These insights highlighted a clear opportunity to reduce information fragmentation and create a more contextual, task-focused repair experience.
These insights highlighted a clear opportunity to reduce information fragmentation and create a more contextual, task-focused repair experience.
Previous work flow
Previous work flow
Previous work flow
What do Geniuses & Technicians say?
What do Geniuses & Technicians say?
What do Geniuses & Technicians say?
"Here's a great little video which is maybe 5sec long, rather than me having to scrub through a video, I don't know how many 3 min or whatever to find a specific part of it".
"Here's a great little video which is maybe 5sec long, rather than me having to scrub through a video, I don't know how many 3 min or whatever to find a specific part of it".
"You're breaking down the barrier between articles — the title of the article no longer be a hard requirement to actually getting the content".
"You're breaking down the barrier between articles — the title of the article no longer be a hard requirement to actually getting the content".
"It's kind of a visual feedback, like click point saying skip-to-this-part — like chapter markers".
"It's kind of a visual feedback, like click point saying skip-to-this-part — like chapter markers".
"Not everybody have the capability to translate the technology to normal, interactive or relatable conversation".
"Not everybody have the capability to translate the technology to normal, interactive or relatable conversation".
"Sometimes I can't get to them — so you have to have a smart little AI assistant to help out in those situations".
"Sometimes I can't get to them — so you have to have a smart little AI assistant to help out in those situations".
"The video for, like, removing the glue on iPhones, it's not really that credible when it comes to the actual real-life interaction with a device that has, like, a one or two year old display".
"The video for, like, removing the glue on iPhones, it's not really that credible when it comes to the actual real-life interaction with a device that has, like, a one or two year old display".
"Ability to save it in the section, if you Bookmark it on your section... That could be really nice".
"Ability to save it in the section, if you Bookmark it on your section... That could be really nice".
Where the Problem lies
Where the Problem lies
Where the Problem lies
Technicians waste time switching between multiple tools and searching through complex documentation, making repairs slower and increasing the risk of missed steps or errors.
Technicians waste time switching between multiple tools and searching through complex documentation, making repairs slower and increasing the risk of missed steps or errors.
Fragmented Access to Knowledge
Fragmented Access to Knowledge
Technicians switch between multiple systems - GSX, Atlas, videos, PDFs - to find a single answer.
Leads to context switching and workflow interruptions
Technicians switch between multiple systems - GSX, Atlas, videos, PDFs - to find a single answer.
Leads to context switching and workflow interruptions
Cognitive Overload During Repails
Cognitive Overload During Repails
Critical data (like warnings, tool specs, or calibration steps) are buried in long-form documents.
Results in missed steps, slower task completion, or over-reliance on memory.
Critical data (like warnings, tool specs, or calibration steps) are buried in long-form documents.
Results in missed steps, slower task completion, or over-reliance on memory.
Inconsistent User Experience Across Tools
Inconsistent User Experience Across Tools
Information is presented differently across interfaces, what's clear in one tool is buried in another.
Reduces efficiency and confidence during live repair.
Information is presented differently across interfaces, what's clear in one tool is buried in another.
Reduces efficiency and confidence during live repair.
Limited Real-Time Support
Limited Real-Time Support
When something unexpected occurs, technicians must escalate or search manually.
Causes delays, downtime, and frustration in high-volume repair environments.
When something unexpected occurs, technicians must escalate or search manually.
Causes delays, downtime, and frustration in high-volume repair environments.
The users we are Designing for?
The users we are Designing for?
The users we are Designing for?
Specialists, Lead Genius, Genius and Technicians
Specialists, Lead Genius, Genius and Technicians
[Click on images to expand]
[Click on images to expand]
Decoding User-stories
Decoding User-stories
Decoding User-stories
Before any wireframes, every interview quote was synthesized into an opportunity matrix, user need, pain point, concept, and UX opportunity per row. Each story produced 2–5 response-pattern options, letting me evaluate the response system before committing to layout.
Before any wireframes, every interview quote was synthesized into an opportunity matrix, user need, pain point, concept, and UX opportunity per row. Each story produced 2–5 response-pattern options, letting me evaluate the response system before committing to layout.
Before any wireframes, every interview quote was synthesized into an opportunity matrix, user need, pain point, concept, and UX opportunity per row. Each story produced 2–5 response-pattern options, letting me evaluate the response system before committing to layout.
Final Conclusion
Final Conclusion
Final Conclusion
"Build an
"Build an
AI Powered, integrated Chat experience
AI Powered, integrated Chat experience
for
for
APPLE Geniuses and Technicians Worldwide
APPLE
APPLE Genius
with a
and Technicians Worldwide
with a
Intent to organize fragmented Articles
Geniuses and Technicians Worldwide
with a
Intent to organize
Intent to organize fragmented Articles and Information across GSX
fragmented Articles and Information across GSX
and Information across GSX
and
and
deliver instant, synthesized answers,
deliver
eliminating context-switching and accelerate decision-making
eliminating context-switching and
in high pressure work flow".
in high
instant, synthesized answers,
pressure work flow".
accelerate decision-making
in high pressure work flow".
Who we are designing for?
Who we are designing for?
What do we want to achieve!
What do we want to achieve!
Where it collects the required information
Where it collects the required information
Why it is absolutely required
Why it is absolutely required
Why it is absolutely required
Integration of Features
Integration of Features
Ideating structural directions
Ideating structural directions
Ideating structural directions
The chosen direction is a hybrid: conversational entry escalating into a video-anchored workspace with an overlay panel system once an active repair begins.
The chosen direction is a hybrid: conversational entry escalating into a video-anchored workspace with an overlay panel system once an active repair begins.
The chosen direction is a hybrid: conversational entry escalating into a video-anchored workspace with an overlay panel system once an active repair begins.
Edge-cases
Edge-cases
Edge-cases
An interface can be perfect and the product still fails if the AI behaves badly at the edges.
Working with the AI/ML engineers on the pod, we specified failure modes — 13 conversational, 2 diagnostic, each with a recommended system behavior, message copy, and lo-fi recovery flows.
An interface can be perfect and the product still fails if the AI behaves badly at the edges.
Working with the AI/ML engineers on the pod, we specified failure modes — 13 conversational, 2 diagnostic, each with a recommended system behavior, message copy, and lo-fi recovery flows.
An interface can be perfect and the product still fails if the AI behaves badly at the edges.
Working with the AI/ML engineers on the pod, we specified failure modes — 13 conversational, 2 diagnostic, each with a recommended system behavior, message copy, and lo-fi recovery flows.
The Skeleton transformation
Protype Glimpse
Protype Glimpse
Protype Glimpse
Free-text stays available the whole time — structure is offered, never forced.
Final Design
Final Design
Final Design
What shipped: a familiar chat shell, with a repair workspace underneath it.
Thank You…
Thank You…
Thank You…
Hope you like case-study
Testimonials
Design Lead
Apple
Tanzil took a highly technical, messy user problem and turned it into a clean, frictionless experience. He delivered ahead of schedule and helped us hit our adoption goals instantly.
Founder
Wordwise Language Labs
WORKING WITH Tanzil WAS SMOOTH FROM START TO FINISH. HE UNDERSTOOD the vission and User NEEDS QUICKLY AND BUILT A SOLID, RELIABLE PRODUCT.
Product Manager
NeoTech
SUPER RELIABLE AND DETAIL-ORIENTED. he can TAKE OWNERSHIP AND SHIP HIGH-QUALITY WORK WITHOUT HAND-HOLDING.
09:34
Testimonials
Design Lead
Apple
Tanzil took a highly technical, messy user problem and turned it into a clean, frictionless experience. He delivered ahead of schedule and helped us hit our adoption goals instantly.
Founder
Wordwise Language Labs
WORKING WITH Tanzil WAS SMOOTH FROM START TO FINISH. HE UNDERSTOOD the vission and User NEEDS QUICKLY AND BUILT A SOLID, RELIABLE PRODUCT.
Product Manager
NeoTech
SUPER RELIABLE AND DETAIL-ORIENTED. he can TAKE OWNERSHIP AND SHIP HIGH-QUALITY WORK WITHOUT HAND-HOLDING.
09:34
Testimonials
Design Lead
Apple
Tanzil took a highly technical, messy user problem and turned it into a clean, frictionless experience. He delivered ahead of schedule and helped us hit our adoption goals instantly.
Founder
Wordwise Language Labs
WORKING WITH Tanzil WAS SMOOTH FROM START TO FINISH. HE UNDERSTOOD the vission and User NEEDS QUICKLY AND BUILT A SOLID, RELIABLE PRODUCT.
Product Manager
NeoTech
SUPER RELIABLE AND DETAIL-ORIENTED. he can TAKE OWNERSHIP AND SHIP HIGH-QUALITY WORK WITHOUT HAND-HOLDING.
09:34