Qpoint Design

Pages

Data Swarm — Ideas

Source: c62 visibility-stream — real-data pairing pass: every mark is a datum, AI expenditure as a first-class progress metric

Four plots where every mark maps to a real event — no decoration. The AI data is this machine's actual Claude transcripts (/api/ai-usage: 54.6M output tokens, 157 sessions); code data is per-commit --numstat (/api/commits/detail: 768 commits). Categorical palette validated (CVD + contrast) per the dataviz procedure.

The thesis, demonstrated on ourselves: visibility allows the controls and oversight that let AI be deployed safely and efficiently. These charts exist because our AI usage is observable — with qcontrol deployed, the same picture comes from llm.usage events for your whole org.

54.6M

AI output tokens

14.6B

context tokens re-read (cache)

157

AI sessions

71k

tokens per commit

5.0M

tokens per shipped story

45 days · one seat of the team (this machine's transcripts) · the team-wide version of this band is the product

Plot 1

The expenditure swarm

x = time · y = output tokens per session (log) · one dot = one real Claude session · size = messages · color = project. The axis choice makes effort visible at session grain: bursts, marathon sessions, quiet days — the swarm is the impression.

qcontroldesignlayerwwwother

157 sessions · hover any dot for project, date, tokens, messages, model · dashed verticals are digest Thursdays

Plot 2

Tokens in → code out

x = a day's output tokens (log) · y = that day's lines changed (log) · one dot = one real day, darker = more recent. The guides are tokens-per-line isoquants — the chart asks the question companies should be asking: what does a token buy?

27 days where both AI and code moved · single sequential hue, darker = more recent · a drift toward the lower-right guide over time = improving leverage

Plot 3

The commit swarm

x = time · y = lines changed per commit (log) · one dot = one real commit · color = repo. Commit grain shows what day grain hides: the #639 cutover towers at ~52k lines, release trains cluster, the Aug 6 batch day reads as a vertical wall.

qcontroldesignlayerwww

768 commits · +462k / −99k lines · hover for repo, date, size, subject · ✓ callouts are digest stories anchored where their evidence chips match commit subjects — derived, never hand-written

Plot 4

The punchcard of attention

x = day · y = hour of day · one dot = one real active hour · size = tokens generated that hour. When does the AI half of the team work? The swarm answers instantly: shaded columns are weekends.

180 active hours in 45 days (UTC) · size = output tokens that hour · one seat — the team-wide punchcard is an llm.usage query · deliberately unannotated, as is the leverage plot: not every visualization needs the meaning layer

Coda

Why these axes

Time is on x wherever the story is rhythm (1, 3, 4) because partners evaluate a team by its pulse, not its totals. Plot 2 removes time from the axes on purpose — leverage is a relationship, so both axes are quantities and time becomes color. Everything log-scaled: real work is lognormal, and a linear axis would let the #639 monster flatten every other mark to invisibility.

The salivation pitch, stated plainly: AI expenditure is a progress metric you can put on the wall next to commits — sessions, tokens, tokens-per-line, hours-of-attention. We can chart ours because we can see ours. Most companies can't see theirs. That's the product.