Nobody could
use our ML engine.
A powerful ML engine. Low adoption. Twelve to fifteen clicks to train one model — data-flow canvases, cascading menus, a hidden play button. Business users couldn't touch it. I owned the redesign end to end. The explainability popup shipped in 9.2; the full workflow revamp is built and test-complete, awaiting release.
The liability I inherited.
The first thing I saw was that red exclamation — “No results generated. Click play to train the model.” The first thing I said: remove that.
The legacy workflow, on the public record.
A tiny request. I turned it into the ML Functions revamp.
Our Principal Data Scientist showed me a screenshot from another tool: “I want this in WebFOCUS.” A simple explainability popup. I could have designed the one component and handed it back.
Instead I dug into why the popup was needed — how model predictions work, what feature importances actually mean. He was glad I took the effort to understand, and that curiosity earned me the trust that grew into the ML Functions / DSML workflow revamp.
The side project that started it all.
The side project that started it all — double-click a prediction to open Explainability: the XWIN impact plot shows why, and you can change a feature's value and recompute to watch the outcome move. It shipped in 9.2, and earned me the trust that grew into the ML Functions / DSML workflow revamp.
The world it lived in.
The old way: to touch one model's hyperparameters you right-clicked the pill and dug through three nested menus — then did it again for each of 4–6 algorithms. Getting to a trained model at all took 12–15 clicks across a data-flow canvas, a spawned browser tab, and a fake “no results” screen that just meant “click Play.”
The world it lived in — everything interconnected.
“Are you trying
to make this dumb?”
The principal data scientist's question. No. I wasn't removing sophistication — I was making it legible. One path for the novice and the expert. Not two modes.
I couldn't fake the domain. So I learned it.
To bridge the knowledge gap, I ran weekly deep-dives with the data scientist and took an MIT certification in Product Design for AI & ML as an accelerator — I paid for it myself. Together they let me have smarter, faster, highly technical conversations with engineering.
I validated my understanding constantly. I was afraid of making something look right that technically meant the wrong thing — a fake win.
The answer wasn't two experiences — a “simple” mode and an “advanced” one. It was one path with contextual education at every step: definitions, tooltips, explanations. Not a dumbed-down UI — an educated one. The sophistication was there; it just waited for you to ask for it.
One path. Dual purpose.
Business analyst
Zero ML background. Needs to get through it and trust the result.
Data scientist
Wants depth and control — and not to feel spoon-fed.
We didn't have two experiences. We had one path — clear enough for someone who'd never touched a model, but not so easy the data scientist felt spoon-fed. The sophistication was there; it just waited for you to ask.
Mapping the full workflow.
Before a single screen: every model type reconciled into one architecture, and the whole flow diagrammed end to end.
It started as wireframes.
Structure worked out at low fidelity — before a single pixel of the real UI.
The result — models you can actually compare.
The Predict Data landing that didn't exist before — pick a dataset, switch between Train and Run, and read models as scannable tiles (score, algorithm, one-click run) instead of right-clicking a filename to open a properties panel. Learn one, you've learned all three.
“Why don't we have
a landing page here?”
There was no entry point to the system at all. I built the Predict Data hub from scratch — and put it where every WebFOCUS user already lived: the right-click menu.
The landing page that didn't exist — so I built it.
The Predict Data landing that didn't exist before — pick a dataset, switch between Train and Run, and read models as scannable tiles (score, algorithm, one-click run) instead of right-clicking a filename to open a properties panel. Learn one, you've learned all three.
An entry point, a split, and a pattern that spread.
No Predict Data landing page existed — so I designed one from scratch. I first tried combining Train and Run on one display; engineering pushed back (too complex to maintain), so we split them into two tabs — a negotiation that balanced UX ambition with technical reality. Dense model tables made comparison impossible for non-experts, so I replaced them with scannable model cards.
The biggest adoption move was the smallest: in WebFOCUS, right-clicking is religion — everything is right-clickable. Putting “Predict Data” in the right-click menu capitalized on the most common behavior in the whole product. It worked so well the pattern was later extended to “Generate Insights” and “Ask a Question.”
And you reach it the way WebFOCUS users already work — right-click.
Right-click any dataset → Predict Data. Right-click is religion in WebFOCUS, so the entry point rode the most common gesture in the product — 6 clicks down to 2. The same move later carried Generate Insights and Ask a Question.
The reach, before and after.
- 1Click + → New
- 2Open Data Flow
- 3A new browser tab spawns (Reporting Server)
- 4Drag the dataset onto the canvas
- 5Switch to the Train Models tab
- 6Drag "Binary Classification" in
- 7Fake "No results generated" → hunt for Play in the top nav
- 8Right-click the pill → 3 nested menus for hyperparameters×6 · per algorithm
Reaching a trained model the old way was a gauntlet — a data-flow canvas, a spawned browser tab, a fake “no results” screen, cascading menus once per algorithm — 12–15 clicks in all. The new entry rides the gesture every WebFOCUS user already knew: right-click the dataset → Predict Data. Two clicks.
The hardest screen
in the revamp.
The confusion matrix. Two charts, a threshold slider, a 4×3 table. A wall of noise at first. Dozens of sessions with the data scientist — paper, pen, live sketching — until it was undeniable.
Four steps in — the guided wizard that leads to it.
Four steps, no branching — Problem type → Target → Predictors → Hyperparameters. Every step carries plain-language help, and defaults are pre-set, so a business user goes from dataset to trained model without a data-science degree. The structure came straight from mapping the domain with the data scientist.
The redesigned workflow, in motion.
Four steps in. One screen that finally made it obvious.
The four-step spine practically designed itself once I understood the domain: Problem Type → Target → Predictors → Hyperparameters. Four things you need to train a model — no branching, no hidden paths.
I tried full-screen views, side-step layouts, two columns. The popup wizard won — WebFOCUS engineers lived in modals, it left room for helper text and definitions, it let each step stand cleanly, and it was faster to build. The best design move is the one that solves the problem without asking the legacy product to become something it isn't.
The confusion matrix was the hard part: a wall of noise when I first saw it. After dozens of sessions I found the move — I aligned the threshold slider exactly 90° parallel, directly under the chart point that moves. Drag the slider, the dot on the line chart moves with it. No way to miss the relationship. Once I understood every technical interaction, I could finally explain it visually.
The real screens — full fidelity.
Beyond the recreations: the shipping product itself, still unreleased.
Comparing models — from a wall of numbers to cards.
The old results were just a list of model names — no scores in sight, so to compare them you opened each one on its own. I replaced it with scannable model cards: problem type, algorithm, and one score you actually read. Compare at a glance, pick the best, hit run.
The confusion matrix — the slider at 90°.
“The best screen in the entire UX revamp,” the principal data scientist said — out loud, in front of everyone.
| n = 2,000 | Predicted: N | Predicted: P |
|---|---|---|
| Actual: N | TN = 1,811 | FP = 35 |
| Actual: P | FN = 58 | TP = 96 |
| n = 2,000 | Predicted: N | Predicted: P |
|---|---|---|
| Actual: N | TN = 1,811 | FP = 35 |
| Actual: P | FN = 58 | TP = 96 |
The principles it ran on.
Explainability first
If I can explain it visually, anyone can understand it — so I didn't design a screen until I understood every interaction on it.
One path, dual purpose
Not two experiences. One flow, legible to a novice and useful to a data scientist at the same time.
Contextual education
Definitions and tooltips at every step. An educated UI, not a dumbed-down one.
Evidence over instinct
“I designed this because I felt like it” means nothing. Every decision was backed by research, existing behavior, or documentation.
From 12–15 clicks
to as few as 6.
Reaching training alone dropped from six clicks to two — right-click a dataset → Predict Data. A full trained model went from 12–15 clicks to as few as six, depending on how much you configure. The always-on data-flow canvas became a guided four-step wizard, and the fake “no results” screen disappeared entirely.
The step-4 breakthrough — every algorithm, one screen.
The old way meant a separate popup for every algorithm, one at a time. My redesign puts a horizontal tab per algorithm into step 4 of the wizard — tune them all in one screen, then one click to Apply and Train Models.
I defended it with data,
not opinions.
The data scientist wanted the old Data Flow canvas back in the UI. I didn't argue from taste. I brought user research and session recordings, and we kept the new direction.
The whole platform, in my head — and the evidence to defend it.
Nobody handed me the map. I built a mental model of WebFOCUS as one end-to-end ecosystem — get data, prepare it, train ML, visualize, distribute — by touching every corner of the product.
I was the only designer on ML — and on IQ, and on the Reporting Server — running all of them alongside ReportCaster. After the layoffs, it was several systems at once.
So when the data scientist tried to bring the Data Flow canvas back into the UI, I didn't argue from taste. In a room with two engineering directors and the principal data scientist, I used user research and session recordings to show why reverting would break the workflow. We kept the new direction.
In testing, they just
blazed through it.
In usability testing, one instruction — “It starts from the dataset” — was enough: four SMEs right-clicked, hit Predict Data, and blazed through the four-step wizard with no help. My design director, zero ML background, trained a model. I taught him machine learning through the UI. (This is validation of the built redesign, ahead of release — not production adoption.)
“The best screen in the entire UX revamp.”
| n = 2,000 | Predicted: N | Predicted: P |
|---|---|---|
| Actual: N | TN = 1,811 | FP = 35 |
| Actual: P | FN = 58 | TP = 96 |
| n = 2,000 | Predicted: N | Predicted: P |
|---|---|---|
| Actual: N | TN = 1,811 | FP = 35 |
| Actual: P | FN = 58 | TP = 96 |
The confusion matrix — the screen he was talking about.
Validated by the people who'd know.
“The clarity of her designs, in spite of the underlying data science and machine learning complexity, is impressive and has greatly contributed to the success of our products.”
“Her design was clean, intuitive, and clearly addressed the needs of users across different skill levels.”
What I'd push harder forLive sessions from week one
Sketching together inside my design tool on video calls — paper, pen, live — was the breakthrough workflow. Starting it at week one instead of halfway through would have saved sprint cycles.
What I'd do nextSimplify even further
Gray out incompatible models before users try them, simplify the results screen further, and let users enter run-model mode straight from training — no context switching.
What a mind-bending
project.
ML taught me product partnership — my PM and I co-owned it. Somewhere in there, I stopped decorating a feature. I was shaping the whole product experience.