Explore Data
in WebFOCUS.
Explore Data is where WebFOCUS answers questions about your data. Two AI features live there: ask a question in plain English with Natural Language Query, and turn a dataset into findings automatically with Insights. Both engines were genuinely powerful. Both sat buried in a popup nobody opened. I designed and modernized each — then helped bring them into one hub.
A command line,
dressed as a web app.
Our Natural Language Query engine was genuinely powerful — ask a question in plain English, get SQL against the reporting server. But the interface was a blank prompt with no first move, and cryptic errors that leaked straight to the user. Adoption was very low. I redesigned it into a conversation.
Powerful engine. Invisible feature.
The technology already worked. The interaction didn't.
The hard part was already solved. A question like “What were the sales for 2019 in March?” was translated into a real SQL query against the Reporting Server — later powered by a Phi-3 Small Language Model. The intelligence was there. What wasn't there was any reason for a business user to trust it, or even to start.
You landed on an empty box. Nothing suggested what you could ask. If the model misfired, you got a raw SQL stack trace or a red error screen. It behaved like a command line wearing a web UI, so the people it was built for never touched it. My job wasn't to add capability. It was to make the capability legible.
What the old UI actually showed the user.
A response “details” dump: the raw SQL, a 500, and “Cannot convert SQL to FOCUS” — handed straight to a business analyst.
Give people
a first move.
The fastest way to kill a conversational tool is a blank page. So the redesign starts by answering the question the user hasn't asked yet — and when things go wrong, it talks back like a human, not a database.
Two moves that turned a prompt into a conversation.
First, I killed the blank page with suggested questions. Land on the tool and it already offers real, answerable questions — ask one, type your own, or shuffle for a new set. There's always a first move, so nobody stalls at an empty box.
Second, errors. When the SLM met a query it couldn't map — the classic “Why is the sky blue?” — engineering's instinct was to surface the raw failure. I designed a Fallback Handler and guardrail instead: low-confidence queries are caught before the SQL call is ever made, and answered with a friendly empty state that offers a way forward. I also partnered with our principal data scientist on the hardware reality behind it — the SLM crawled to 30-minute lag on CPU-only servers, so we set an engineering policy that IQ requires and validates GPU-enabled server profiles, alerting the admin through the Hub when it's running on CPU.
Suggested questions — the blank page, solved.
Ask your question above.
A blank prompt is where adoption goes to die. Suggested questions hand you a first move — pick one, type your own, or hit Refresh suggestions for a new set — so nobody ever faces an empty box.
The same idea, shipped — at full fidelity.
Conversational guidance in place of an empty prompt — the shipped empty state.
And when there's no answer, it says so — kindly.
A nonsense question used to leak a raw response trace — 500 · Cannot convert SQL to FOCUS — straight to the user. The guardrail catches low-confidence queries and answers like a person: it points you back to suggestions it can actually answer, never a stack trace.
The same screen, redesigned — the shipped “after.”
Where the old UI dumped raw SQL, the redesign shows a human message: “We're having trouble connecting to the data source.”
One answer,
read every way.
The move that was mine end to end: a chart switcher that lets a single answer be re-read as a table, bar, column, line, or area. I conceptualized it and got PM and leadership behind it — and it grew into far more than NLQ needed.
The chart switcher — my idea, shipped small.
One answer, five ways to read it — table, bar, column, line, area — from a single toolbar, with the stacked column as the shipped default. This mini switcher is the tip of a bigger idea I conceptualized and sold to PM and leadership: a multi-use chart component spanning 36+ types for both Designer and NLQ. That full component was built but never released — the switcher is what made it into 9.3.
An idea I sold up the chain.
The switcher started as a small control — one answer, re-read as a table, bar, column, line, or area — but I saw it could be bigger. What I conceptualized for NLQ grew into a multi-use component of 36+ chart types meant to serve both Designer and NLQ across the platform. That full component was built but never released; the disciplined, mini version is what shipped inside NLQ in 9.3 — the restrained slice that earned its place in the release.
The principles it ran on.
Never a blank page
The tool always offers a first move. An empty prompt is a dead end; a suggestion is an invitation.
Fail like a human
Catch the bad query before the model does, and answer with a way forward — never a stack trace.
Ship the disciplined version
The 36-chart component was the vision; the mini switcher was what earned its place in the release.
+25% adoption.
Live in 9.3.
NLQ was re-architected from a SQL-based translator to Microsoft's Phi-3 Small Language Model. The engine change (SQL → Phi-3) and the redesign built around the new model's capabilities together drove a ~25% jump in adoption. My redesign is what made the faster engine legible — a first move, human errors, one answer read every way. It ships in WebFOCUS 9.3 today, still inside the Explore Data popup; whenever the IQ Plugin launches, its visibility grows again.
Shipped, live, and validated by the people who built it with me.
“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.”
“Anuja demonstrated exceptional ability to understand intricate workflows and translate them into elegant, user-centric designs that elevated the product’s usability and visual appeal.”
Where to see itLive in WebFOCUS 9.3
NLQ shipped in 9.3; the modernized experience appears in ibi's public WebFOCUS 9.3 material.
What I'd do nextRelease the full switcher
The 36-chart multi-use component is already built. Finishing and releasing it would let a single answer be re-read any way the analyst needs — across both Designer and NLQ — instead of the five the mini switcher ships today.
You had the answers.
You couldn't see them.
WebFOCUS already ran a powerful automated-insights engine. But every finding sat collapsed in a long list of accordions — you opened each one by hand to read it. So nobody used it. I turned it into one move: pick a dataset, hit Generate, read the findings in plain language.
The engine worked. The interface buried it.
The intelligence was real — signals, trends, correlations, computed automatically. But the output was a long list of collapsed accordions — you clicked each one open to find out what it said, with no ranking and no plain-language read. A business user opened it once and left.
A grid answers nothing. It just shows you everything.
Insights sat inside the Data Science & ML suite, next to the model tools. It could scan a dataset and surface what stood out — no query, no chart-building. Technically, a lot of value for a single click.
But the result landed as a long list of accordions — every insight collapsed, so you opened each one by hand to see what it said, with no ranking and no plain-language read. To get anything out of it you had to already know what you were looking for. The people it was built for — analysts who didn't — bounced off it. A powerful feature, invisible in plain sight.
It starts from
the dataset.
Three moves, no query language: pick a dataset, confirm the fact table (auto-detected, or choose your own), hit Generate Insights. Seconds later — findings, not a spreadsheet. And you reach it the way WebFOCUS users already work: right-click.
The whole feature is three moves — dataset → fact table → Generate Insights. No query to write, no chart to configure. Reached the way WebFOCUS users already work: right-click a dataset. Seconds later, plain-language findings, not a raw grid.
Design the path in. Then design what greets you when there's nothing yet.
The entry point rode the most common gesture in the whole product: in WebFOCUS, right-clicking is religion — everything is right-clickable. The plus-menu options already existed, but grouping Predict Data, Generate Insights, and Ask a Question onto a dataset's right-click menu was my suggestion — a shortcut that became even more useful with the later IQ plugin.
From there the flow is deliberately flat — dataset, fact table, generate — with a sensible default at every step so a first-timer never stalls. And because the first thing many users see is an empty panel, I designed that too: a friendly illustration, a plain statement of what you'll get, and one obvious action. No dead ends.
The empty state, doing real work.
Even the empty state does work. Instead of a dead panel, its own friendly illustration and one obvious action tell you exactly what to do next to get your first findings — no loose ends.
The principles it ran on.
Read, don't decode
A finding is a sentence, not a row of numbers. The narrative leads; the chart is there to back it up.
Scannable over complete
Tiles you can sweep in a glance beat a list of accordions you have to open one by one.
Meet the existing behavior
Right-click was already the muscle memory. The entry point rode it instead of fighting it.
No loose ends
Even with zero data, the screen tells you exactly what to do next and what you'll get for it.
Findings you can scan,
not decode.
The engine already found these. The redesign gave them a face — each finding a tile with a plain-language headline and just the chart that backs it up. Its own color palette, tuned so every category stays legible at a glance.
The engine found the same insights either way. Before, they sat collapsed in a long accordion list — you opened each one to read it. I turned every finding into a scannable tile: a plain-language headline and the chart that backs it up, readable at a glance.
And here it is, shipped.
The real Insights results in WebFOCUS — the same scannable tiles at full fidelity: each finding a plain-language headline with just the chart that backs it up.
Insights got its own colour palette — and rules to go with it.
A data scientist pointed out that the Insights charts didn't follow a coherent palette and didn't match Designer's. I built a dedicated Insights palette and sample charts to demonstrate it — a theme that could be saved and reopened in Designer for further customization.
Then I validated it across every chart type.
The palette wasn't just colours — it was a set of rules, then tested across bar, ring, scatter, line-with-trend, density, and correlation-matrix charts, so a finding reads clearly no matter how the engine chooses to visualize it.
Down to the correlation matrix — a palette specced by value.
The sequential ramp and the both-ends diverging scale for correlation charts, each step pinned to an exact RGB value so a matrix reads cleanly from +1 through 0 to −1.
And the type, specced to the point.
Every label on an insight chart — title, axis titles and labels, legend, footer — pinned to a size, a weight, and a token, so a finding reads the same no matter which chart the engine picks.
And a tracker for every insight running.
Generation can take time, so I designed an Insight Generation Status view — every insight, its dataset and fact table, and its state (Ready, Queued, In Progress, Error, Deleted). In the shipped Explore Data version it's a modal (and a new-browser-tab generation view); the IQ side-panel version is design that hasn't shipped.
It shipped. It's live.
Generate Insights went out in the Data Science & ML suite and is live in WebFOCUS 9.3.
Validated by the people closest to the data.
The engines didn't change.
The conversation did.
Neither feature was a new-capability project. Both engines already shipped and sat idle — the intelligence was there, but people couldn't reach it. The work was two decisions about what “asking your data” and “a result” should feel like: a first move instead of a blank box, a sentence instead of an accordion. Once Explore Data felt human, people finally used what we'd built all along — and that's the ground the IQ hub was built on.