From manual to conversational
Desiginig the
Assistant
May 2024
What can I help with?
Analyze data
Explore a dataset
Generate data
From Drive
My Role
Head of Product Design
Team
Duygu Okcu, PO/PM Mario Scriminaci, CPO Michael Platzer, CTO Iñaki Guastali, SWE Giorgi Bakradze, SWE Kuba Bielawski, SWE Zack Tilev, UXW Artur Mikłasewicz, QA
Timeline & status
2 months
Overview
MOSTLY AI faced a critical challenge in data democratization: while the platform could unlock sensitive data through synthetic generation, accessing and understanding that data remained limited to technical users. Complex configuration screens created barriers, and non-technical teams had no way to explore data or gain insights.
I led the design of the MOSTLY AI Assistant — a conversational AI interface that transforms how users interact with data. Through natural language, users can now analyze data, uncover insights, generate synthetic datasets, and configure generators without touching a single configuration screen.
This shifted MOSTLY AI from a technical platform for specialists to one that truly democratizes data — making it accessible and actionable for everyone. The Assistant opened entirely new use cases while dramatically reducing time-to-value.
Assistant Capabilities
Data analytics & insights
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
Ask questions about your data and get instant answers, visualizations, and insights.
CONTEXT
Powerful capabilities, but only accessible to technical specialists.
Data locked behind technical barriers
Despite offering powerful synthetic data generation, MOSTLY AI faced a fundamental adoption challenge. Users needed technical expertise to configure generators and couldn't extract insights without Python or data science skills.
This created a paradox: we could unlock sensitive data through privacy-safe synthetic generation, but only technical specialists could use it. Business analysts, product managers, and domain experts–the people who often understood the data's context best–were locked out, preventing the very democratization the platform was designed to enable.
1.0 Legacy end-to-end workflow
1.1 Direct insights from enterprise users
1.2 Direct insights from free-tier users
Systemic friction, clearly felt at every level
Whether enterprise or free users, everyone ran into the same underlying pain: a rigid system that resisted iteration, clarity, and control.
It didn’t matter if you were a data scientist in a bank or a developer testing the free version — the experience felt heavy, slow, and unforgiving. The core architecture made simple tasks feel complex, and complex tasks nearly impossible.
Connectors were one-way: you could only read data, with no option to deliver synthetic data back.
Catalogs acted like templates but lacked clarity, forcing users to manage everything in one place.
Jobs coupled training and generation, forcing full retraining even for minor output changes.
There was no flexibility to iterate — tweaking one thing meant restarting everything.
All users were pushed through the same complex flow, regardless of technical skill or intent.
THE CHALLENGE
Take a tightly coupled generation pipeline and make it modular, reusable, and simple for everyone.
IMPACT
A product that finally matched the ambition behind it.
Momentum by design
This initiative didn’t just refine the experience — it redefined how we work, aligned the product with user needs, and gave the team the confidence to move fast without breaking things.
40% faster time-to-first dataset, streamlining workflows and reducing friction for data consumers.
Model reuse unlocked, allowing users to train once and generate synthetic data on demand.
Async workflows enabled, so users could configure datasets without waiting for model training to finish.
Dual-role support unlocked, enabling model creators to focus on training while data consumers generate datasets independently.
Release cadence increased nearly 8×, from one release every 68 days to one every 7.6, with 30 modular releases shipped over 10 months.
Unlocked strategic momentum, earning trust from leadership and opening up new opportunities for the team.
Faster time-to-value for new users
Redesigning the experience dramatically reduced the time it took new users to generate and download their first synthetic dataset — from over 5 minutes to just above 3. This 40% improvement reflects a more intuitive flow, fewer barriers, and a quicker path to meaningful outcomes.
2.0 Faster path to first synthetic dataset
From perfectionism to momentum
The redesign wasn’t just visual — it redefined how we build. We shifted from slow, monolithic releases to a fast, feedback-driven cadence. By breaking work into smaller, shippable units, we released updates 8× more frequently, putting ideas in front of users sooner and learning what truly mattered — before over-investing.
2.1 Release cadence shift over time
HIGHLIGHTS
A full-stack product evolution that brought clarity, flexibility, and scalability to synthetic data workflows.
2.0 Product homepage evolution
2.1 Generator configuration screen sample UI
2.2 Generator entity overview page
2.3 Synthetic dataset configuration page
THE PROBLEM
When everything blocks everything
A pile-up of constraints
Designing wasn’t the only challenge — the product had to serve both enterprise teams and free users, each with very different needs.
But with no structured research beyond user interviews, it was hard to ground decisions in real insight. Internally, design specs lacked clarity, the system in place struggled to scale with product demands, and collaboration across teams often left intent behind.
All of this, under a tight deadline, made it clear: fixing the product meant fixing how we worked.
A limiting design system struggled to keep up with the product's pace, maturity and evolving needs.
One product, two audiences required balancing enterprise needs with self-serve simplicity in every design decision.
A tight two-month deadline meant the redesign had to move fast and launch in stages.
Collaboration fell short — specs lacked clarity, design intent got lost, and alignment across teams was fragile.
Research was ad hoc — beyond user interviews, there were no structured ways to gather or apply insights.
FOUNDATIONAL GAPS
A rigid system, unclear processes, and misaligned teams — we weren’t just overdue for change, we were held back by staying the same.
A BETTER FLOW
The architecture that unlocked growth.
You gotta start somewhere.
The redesign introduced reusable Generators, standalone Synthetic Datasets, and bidirectional Connectors — decoupling training from generation and unlocking true flexibility.
It replaced the rigid, overwhelming flow of the old interface (Fig. 3.0) with something clearer, more modular, and easier to use — enabling optionality and speed at every step (Fig. 3.1).
3.0 Legacy interface — overwhelming, rigid, and hard to scale
A flexible foundation
Generators can now be reused: train once, generate endlessly. Synthetic Datasets are configured independently, and data can be delivered anywhere.
The modular architecture supports a flexible path through the product — introducing a Zero Config Mode for new users who want quick results without heavy setup, while still offering advanced controls for experts and enterprise teams.
This approach removes bottlenecks, accelerates iteration, and adapts to different levels of experience.
3.1 Redesigned end-to-end flow
Generators let users train once and generate synthetic data multiple times — no retraining needed.
Independent configurations enable fine-tuning of rebalancing, imputation, and generation mood without touching the model.
Zero Config Mode streamlines the flow for first-time users, skipping complexity and accelerating time-to-value.
On-Demand generation gives users the freedom to synthesize data whenever and however they need it.
Bi-directional Connectors allow not just reading from sources, but delivering synthetic data back to any destinations.
ALIGNMENT & MOMENTUM
Turning a vision into a shared direction
Aligning on the new direction.
Driving a redesign of this scale wasn’t just about pixels – it was about alignment. Before a single component shipped, I presented the new vision at a leadership offsite to the C-suite and extended management team – showing not only what we were changing, but why it mattered. The room’s response validated that we were on the right path.
To scale that energy across the company, I recorded an end-to-end walkthrough of the new product experience. This async artifact helped kick-start conversations, create visibility into the design direction, and unify teams around a shared understanding – from engineering to customer success.
Sharing the redesign at the leadership offsite.
Video walkthrough aligning the company around the redesign.
A SMOOTHER PATH TO SETUP
The configuration that puts users first.
What should’ve been simple, wasn’t.
Before the redesign, configuring data meant navigating fragmented, mandatory flows spread across multiple disconnected screens. Each path was rigid, mutually exclusive, and poorly signposted — turning a basic setup into a source of confusion, delays, and frustration.
4.0 Legacy configuration flow
Pick your level of control.
The redesigned configuration screens made it easier to get started — without sacrificing power. First-time users see a simplified experience, while advanced users unlock granular control when needed.
Smart defaults, optional steps, and inline guidance help users move faster with confidence.
Every choice was tested and validated with real user feedback to reduce friction and build trust.
4.1 Redesigned configuration flow
RESEARCH IN ACTION
The validation that refined every screen
What worked, what didn't.
This study was designed to stress-test the new configuration flow and validate core UX assumptions.
By focusing on real user behavior and expectation alignment, we identified what felt intuitive, what needed simplification, and where friction still remained.
The insights directly shaped the final release.
Methodology
We asked participants to complete 7 tasks spanning the new configuration flow. Tasks were selected to reflect common behaviors and known pain points.
We measured success rate, user expectations, and friction to validate design decisions and uncover areas needing refinement.
Participants
15 Total
Exploration (7)
Evaluating synthetic data, early testing
Production (Single Use Case) (4)
One live workflow using synthetic data
Production (Multiple Use Cases) (4)
Embedded across organization
5.0 Design iterations and prototypes
Clear wins, clearer signals.
The usability study validated the direction — the new designs outperformed the old across all key metrics.
Success rate and expectation matching saw notable improvements, while satisfaction scores hit the ceiling.
The results didn’t just confirm better usability, they built the case for rollout.
5.1 Results summary
GENERATOR OVERVIEW
A hub for synthetic data generation
A clear next step after training
The redesigned Generator overview centralizes everything users need before generating synthetic data. It surfaces key training metrics, accuracy scores, and sample previews, while making configuration and performance data easily accessible.
From here, users can seamlessly jump into generation or analyze model quality – all from a clean, focused view that balances transparency and usability.
6.0 Generator overview
DETAILS MATTER
Thoughtful details that shaped the experience
Small details, big difference
From subtle animations to contextual nudges, every interaction was designed to reduce friction and guide the user with confidence.
By refining spacing, hierarchy, and transitions, we turned moments of hesitation into moments of flow — helping users feel in control without needing to think twice.
Connector
Generators
Synthetic Datasets
7.0 Entity icons
Tracking progress through every stage
Each Generator and Synthetic Dataset moves through a defined lifecycle – from creation to completion. Status pills visualize this journey with precision, helping users instantly understand where an entity stands and what’s happening behind the scenes.
7.1 Entity lifecycle status pill
See the model come to life
This progressive component breaks down each training step with clarity and speed indicators. By visualizing the full pipeline – from data fetch to model report – users know exactly what’s happening and where they are in the process, reinforcing transparency and trust.
7.2 Training status component
Hard to miss, easy to act
The old notification appeared subtly in a cluttered part of the UI and was easy to miss. The new design commands attention with motion, surfaces helpful context, and introduces a clear CTA to drive action.
Daily credit limit reached
7.3 Old notification
7.4 Redesigned notification
A container that adapts
We moved away from the side-drawer pattern, which was tightly coupled to already cluttered screens. The new pop-up provides a cleaner, more focused experience – untethered from the underlying content and flexible across workflows.
7.5 Pop-up surface introduction
Design specs that told the full story
Gone were the days of disconnected screens dropped into Figma. Each spec was now a clear, contextual walkthrough – complete with user goals, flow rationale, interaction logic, and edge-case handling. This shift didn’t just elevate design quality – it sped up implementation, reduced back-and-forth, and brought design and engineering closer together.
7.6 Before – scattered screens, no context
7.7 After – structured, detailed, implementation-ready
OUTCOME
From overhaul to measurable impact
More than a redesign
We didn’t just ship a redesign – we changed how people engaged with the product.
To validate the impact, we isolated the core workflow that the revamp focused on: training Generators and generating Synthetic Datasets.
The plot below shows a clear shift. Right after the first improvements rolled out, we saw usage spike – and it kept climbing. Average monthly growth jumped from 20% before the redesign to 2.4× faster during the rollout period. By year-end, this translated into a 484% increase in active users, all within this flow alone.
8.0 Generator training & Synthetic Data generation flows
A clearer path, a higher finish line
By simplifying the interface and reducing friction across the Generator to Synthetic Dataset flow, we saw completion rates jump from a baseline of 18% to a steady 53%. That’s a 3x improvement in users reaching the end of the journey.
The redesign helped users better understand what to do next, reduced hesitation, and made the system feel more predictable – resulting in fewer drop-offs and more value delivered.
8.1 Task completion uplift after redesign rollout
From frustration to follow-through
Before the redesign, over half of users dropped off mid-flow. With a 52% baseline drop-off rate, we were losing most users before they ever reached value. The redesigned flow drastically improved clarity and guidance, reducing confusion and hesitation.
The result? A steady drop in attrition, eventually stabilizing at just 22% – a 58% reduction. The changes didn’t just look better – they helped more users stick with the journey until the end.
8.2 Drop-off rate before and after redesign rollout
From zero to signal
When I joined, there was no CSAT tracking in place.
I implemented it in my first week to establish a baseline for user sentiment and track how design changes moved the needle over time. What followed was a clear, sustained improvement – validating that the redesign wasn’t just functional, but meaningful to users.
AVERAGE CSAT
3.2
of 5
before redesign
AVERAGE CSAT
4.7
of 5
after redesign
FINAL CHAPTER
A redesign that reshaped more than the product
From delay to direction
This wasn’t just a redesign – it was a shift in how we build, ship, and align. In less than a year, we turned a stalled project into a modular, future-ready platform that grew adoption, improved sentiment, and accelerated decision-making.
Along the way, we uncovered valuable lessons about collaboration, prioritization, and what simplicity truly means – insights that continue to shape how we work.
Clarity beats cleverness
Users moved faster and made fewer mistakes when the UI was explicit, not smart.
Involve cross-functional partners early
Alignment from engineering, product, and customer success unlocked speed and cohesion.
Constraints reveal opportunities
Time, legacy tech, and limited resourcing pushed us to find sharper, more creative solutions.
Design alignment drives momentum
A shared, visualized vision helped teams move fast and implement with confidence.
Simplicity means less complexity, not fewer features
Streamlining flows without removing power made the platform more approachable.
LASTING CHANGE
A cleaner flow, a shared vision, a faster path to value –
not just for users, but for the teams behind it.
© 2025 Alex Ichim. All rights Reserved.
Made with coffee and long nights.
MAIN
Work
About
CONTACT
Resume
Live updates underway