MOSTLY AI’s platform redesign
Transforming a rigid data flow into a modular system that actually scales
3x
Flow completion
−58%
Drop-off rate
4.7/5
Satisfaction
Designing the Assistant — From manual to conversational
Overview
MOSTLY AI could turn sensitive data into safe synthetic data. But only technical specialists could use it. Analysts, PMs, and domain experts, the people who understood the data best, were locked out behind configuration screens and Python.
I led the design of the Assistant from early vision to launch, then through the iteration that made it the platform's primary interface.
Role
Head of Product Design
Duration
Feb 2024 – Dec 2025
V1 shipped Apr 2024
Collaborators
Product Management
Engineering & ML
AI Research
UX Writing, QA
Tools
Figma
Cursor
Claude Code
The problem
Only specialists could use the platform
Every insight needed Python or a data team. Most people just waited.
Users left the moment they got their data
Generate, export, gone. Nothing brought them back.
The bet
One interface, many problems solved
Conversation didn't fix one thing. It collapsed five separate problems into a single bet.
How it works
Conversation, separated from computation
The model handles language, never your data. Real code runs the analysis in an isolated environment, and returns a result you can verify.
Designing the experience
Conversation took the center
The old homepage made you pick a path before you knew what was possible. The new one just asks what you need.
A blank input is a wall
An empty box is intimidating. You don't know what to ask, so the box shows you, cycling through what's possible while the conversation starters adapt to who's asking.
The starters aren't fixed. Broad for newcomers, personal for returning users, surfacing the connectors and actions they reach for most.
Showing what's possible, mid-sentence
As users type, the Assistant reveals what it can do, and surfaces new features the moment they ship.
Frequent for newcomers, off for experts. New features reach everyone, whatever their settings.
The whole job, inside the conversation
Connect a source, ask in plain language, and the Assistant builds, trains, and hands back results you can inspect. No screens, no leaving the chat.
Real platform objects — connector, generator, synthetic table — pulled into the chat, recognizable at a glance.
Show the work, not just the answer
Trust came from transparency, not polish. Every answer unfolds its reasoning step by step, with the real Python one click away, then folds back into an auditable trail once it's done.
Steps stream in as they run — plain-language reasoning, the exact code one click away. When it's done, the trail collapses to keep the chat clean, still there to inspect.
Where conversation wasn't enough
Some work still lived in complex configuration screens. So the Assistant came to those screens, reading the settings, understanding intent, and changing them directly.
The user described a worry in plain words. The Assistant named the right control, turned it on, and the real settings updated in view, each AI change marked.
Never wonder if it's working
Each conversation runs its own Python kernel, which takes time to start and costs real compute. So both were always visible, never a black box.
Design solved latency perception before engineering solved latency. Live status turned dead wait time into visible progress.
What usage revealed
Launch was the start, not the finish. Watching real use surfaced the next round of work.
Sharing, built in
Usage analytics showed people exporting artifacts from the chat one by one to share them outside the platform. So the conversation learned to package itself.
A whole analysis became a shareable report, editable and public, without ever leaving the chat.
Permission before action
Some requests were too consequential to run blind, pulling live external data, combining it with their own, building a report from it.
For high-stakes work, the Assistant proposed a plan first. Nothing ran until the user said go.
Economics that scaled
Heavy adoption pushed compute past projections. The free model that drove growth wasn't sustainable on its own.
A Pro tier protected free exploration while making the economics work.
Outcome
Conversation became the platform
77%
The Assistant became the primary way people used MOSTLY AI.
4.5x
People used to generate their data and leave. Now they stay to analyze and build.
10x
From a handful of technical roles to anyone with a question about data.
The roles once locked out now work through the Assistant — alongside many more it never used to reach.
What I learned
Show the work, don't just deliver it
Transparency built trust faster than accuracy did. Users who could see the reasoning and the code became power users. The ones who couldn't stayed skeptical.
Design for the middle, not the edges
Optimizing for experts shut everyone else out. Optimizing for beginners felt patronizing. The product worked when it served people with intent but incomplete knowledge.
Conversation has limits, and naming them is the job
The bet was conversation everywhere. The truth was that complex configuration still needed real screens. Knowing where not to apply the idea mattered as much as the idea.
The real shift wasn't access to data. It was who got to ask the questions.
