Loupecase studyDashboard →

Dashboard design

What this is for: Fix what the dashboard shows and how a reader interacts with it before the build is judged. Date: 2026-09-20 Status: Final

Who reads it and when

Priya, the PM persona from the PRD, opens the dashboard on a Monday morning with about five minutes and one question in mind: is usage growing, what are people using the assistant for, or where is it failing. She reads one view, gets an answer with its caveat, and closes the tab.

Information hierarchy

Views

Overview. "How much is happening, and is anyone coming back." Six tiles: conversations, turns, weeks covered, peak weekly pseudo-users, return rate for the latest complete week (with its week date), intent coverage. Then the view's insight card asks "Is usage growing, or is it the same power users?" and answers it from the aggregates: conversations over the last 12 complete weeks against the 12 before, the latest top-10% concentration, and which model served recent traffic, with a link to finding F3. Below, weekly conversations as a stacked area by model: the height is the week's total and each color is the share one model family served, so the chatbot's model handoffs read as one color giving way to the next. (The first version drew eight separate lines; because each line began and ended at a handoff, readers saw lines "ending randomly" and asked whether the dataset was incomplete, and a 2026-10-02 review replaced it.) Isolating a model from the legend shows its own area alone. Hover gives a tooltip on every point.

Intensity. "Is growth broad or concentrated, and did people come back." Tiles carry the week they describe, and the return tile names the latest week that has a following week to measure, so it is never blank. The insight card asks "Are people coming back?" and answers with the pre-break median return rate, the collection change that makes later weeks incomparable, and the single-turn share by model, linking to F1, F2 and F4. The return-rate line draws NULL weeks (no following week to measure) as gaps, not interpolated. Every weekly chart on this view carries a dashed vertical rule labeled "collection break" at 2024-10. Pseudo-users and concentration are two aligned panels on one time axis: pseudo-users per week as a filled area above, top-10% share as a line below with a reference line at 50%, so "fewer people, more concentrated" reads at a glance. (The first version encoded the share as the color of a single line; a design review found it unreadable and it was replaced on 2026-10-02.) Conversations per pseudo-user shows the median (solid) and 90th percentile (dashed) in two colors with a legend. A bar chart shows the share of pseudo-users active in more than one week, by quarter, with post-break quarters in a muted tone. Session depth is one horizontal 100% stacked bar per model family, colored by turn bucket on a sequential ramp and sorted by single-turn share; isolating a bucket from the legend ranks the models on it. (The first version used nine small multiples, whose axis labels collided.) Tooltips on every mark; gaps never filled in.

Intent. "What are people trying to do." The insight card asks "What are people using it for, and what is shifting?" and answers with the two largest classes, the biggest riser and faller over the last 12 complete weeks against the 12 before, and the share of other (greetings, tests, gibberish) in recent traffic, linking to F8 to F10. A collapsed reliability card follows: held-out accuracy against the two-rater ceiling in one line, expanding to the gate and per-class F1, so a reader knows which class differences are real before reading a chart. A 100% stacked area shows intent share by week; isolating one class from the legend redraws it as a single share line, which is how a trend in one class is actually read. Two 100% stacked bar sets repeat the same seven classes by model family and by the top ten languages; isolating a class ranks the models or languages on it. Classes keep one color across all three charts, assigned in volume order.

Friction. "Where is the assistant failing people." The insight card asks "Where does the assistant fail people most?" and answers with the worst refusal cell among intent × model pairs with at least 1,000 conversations, the one-and-done trend over the last 12 complete weeks, and which real-work intents draw the most single-turn exits and refusals, linking to F5 and F11. Four proxy rates over time (repeated request, correction follow-up, refusal, one-and-done) run as four lines with a clickable legend, on a percent-formatted axis that autoscales to the data. A heatmap has the four signals on one axis and intent-and-model rows on the other, one sequential hue for the rate, never doubling as identity. A table below gives each intent's count and the same four rates with all models combined, so a striking percentage over a small denominator cannot be over-read. A proxy failing its 70% precision gate is dropped outright, not softened; with intent unlabeled, the heatmap and table fall back to "not available" while the weekly lines, which need no intent, still render.

Data quality. "How much to trust everything else." Tiles: minimum cell size (20), shards processed (86 of 86), taxonomy version, aggregate set size. The insight card asks "What in these numbers can I trust?" and answers with intent label coverage and classifier accuracy against the rater ceiling, the share of conversations without a usable country, and token-field coverage, linking to F6 and F7. Below: redaction and empty-input rate as two lines with a legend, on a percent-formatted axis; token-usage coverage over time, labeled with the window it covers from 2024-09-09; top-countries and top-languages bars, each with two muted residual bars, not recorded and small cells, named in the card note so a reader sees how much geography is missing versus privacy-suppressed.

Query. Whatever the other five views did not anticipate. A SQL textarea runs against the twelve committed aggregate views, capped at 200 rows, with a visible error line for a bad query. No chart here; it is the escape valve.

Wireframes

At 400 px, tab pills wrap onto two or three rows, not a menu; tables scroll inside their card; charts scale to the card width.

Overview, desktop.

+---------------------------------------------+
| Loupe   [Overview][Intensity]...[Query]      |
| banner                                       |
+---------------------------------------------+
| [Convs][Turns][Weeks][Peak][Return][Intent]  |
+---------------------------------------------+
| weekly-convs-by-model line, legend           |
+---------------------------------------------+
| caveats, ODC-By                              |
+---------------------------------------------+

Overview, 400 px.

+----------------+
| Loupe          |
| [Ov][Int][Fr]  |
| [DQ][In][Qry]  |
| banner         |
+----------------+
| [Convs][Turns] |
| [Weeks][Peak]  |
| [Return][Int]  |
+----------------+
| chart, scaled  |
+----------------+
| caveats        |
+----------------+

Friction, desktop.

+---------------------------------------------+
| Loupe   [Overview]...[Friction]...[Query]    |
| banner                                       |
+---------------------------------------------+
| proxy-rates-over-time lines, legend          |
+---------------------------------------------+
| heatmap: x=4-signals, y=intent-model, ramp   |
+---------------------------------------------+
| table: intent, n, 4-rates (models-combined)  |
+---------------------------------------------+
| caveats, ODC-By                              |
+---------------------------------------------+

Friction, 400 px.

+----------------+
| Loupe          |
| [Ov][Int][Fr]  |
| [DQ][In][Qry]  |
| banner         |
+----------------+
| rates, legend  |
+----------------+
| heatmap scroll |
+----------------+
| table scroll   |
+----------------+
| caveats        |
+----------------+

Interaction rules

Tooltips appear on every mark, on hover and on tap. Every chart has its own legend, drawn by the app rather than the plotting library so it can act as a control: clicking a series isolates it, shift-clicking adds or removes series to compare, hovering a chip dims the others, and "Show all" restores the chart. Isolation is per chart. Above the views, a model picker ("All models" plus one chip per model family, ordered by volume) filters every view at once: tiles, insight cards, so-what lines and charts re-render for that model, a banner names the model and its conversation count with a one-click reset, and the choice lives in the URL hash (#intensity/gpt-4o) so a filtered view can be shared. The filter reads four weekly aggregates re-cut by model under the same suppression rule; return becomes same-model retention and the tile and definition say so. Charts with no per-model cut (countries, languages, quarterly persistence, intent by language) stay whole and carry an "All models" tag instead of pretending to be filtered. Models with under 24 complete weeks get insight copy that says there is no earlier window to compare against rather than a meaningless delta. Weekly charts follow one convention for missing data, stated in a key under each figure. A count over a week with no conversations is a true zero and is drawn as one, so a model that was switched off drops to the baseline instead of being bridged by a straight line. A rate over no conversations is undefined, so the solid line stops and a short dotted segment carries the eye across to the next measured week; the tooltip says why. The three weeks in which the whole collection was offline (Jan 13, Jan 20 and May 26, 2025) are shaded on every weekly chart, and so are a selected model's own zero-traffic weeks, with the dashed collection-break rule in the same muted tone. Lines are never broken outright and never silently interpolated. (The first version let the plotting library connect across missing weeks, which drew gpt-4 at steady volume through a summer in which it served nothing; a design review on 2026-10-02 replaced it.) When a model is selected, the banner states its window in the logs and its weeks with traffic, the time axis fits that window, and the volume chart marks the last week in the logs, so a retired model reads as retired rather than as missing data. Each chart title is followed by a one-line "so what" computed from the data (latest value, change against the prior 12 weeks, or the ranking the chart shows); the definition of the measure moves under a "What this measures" disclosure so it is available without competing with the reading. The partial final week is excluded from every weekly chart and tile. Nothing animates: a static build regenerated by one batch command, at most weekly, so motion would imply a cadence the pipeline does not have. Color comes from the palette tokens. Rates use percent-formatted axes that autoscale to the data. Suppression is noted in the card note wherever min_cell applies. The suppressed_or_unknown row is drawn in a muted tone and named in the card note. Intent colors are fixed per class across the three intent charts.

Deliberately absent