Token Dashboard
A team token-cost dashboard for engineering managers at a top-5 US bank (what is my team spending on AI, against what budget, on what work), built as a five-day MVP before the formal review process existed. Reading it now, it's the prototype the process was abstracted from.
- 5-day MVP
- Enterprise client
- Process prototype
Origin
A leadership sync on June 9, 2026, plus a client one-pager sent the same day. The client's homegrown token-tracking script "didn't work… not accurate at all," and the deadline drove everything: a functional, rock-solid tracking MVP in five days, dashboards included, even without high-end polish. The single primary user is the engineering manager; developers explicitly don't use this UI; their counters, warnings, and halts live in the IDE and roll up here.
The prototype the process grew out of
This stream predates the numbered pipeline entirely: no objectives doc, no data-relevance doc, no persona audit. Yet the wireframe already does, informally, everything the later steps made explicit: it names its single primary user, tags every panel to a source quote, brackets everything unknown, and refuses to fabricate. What it lacks (a separated data-grounding step) is precisely the gap the later streams' process filled. That arc, prototype to formalized pipeline, is the story of the whole section.
Design decisions
Every number is a door.Three global interaction rules: every engineer name links to a profile, every PR links into the client's source control, every KPI carries a trend.
Refuse "spend avoided." The figure is uncomputable once a session is halted, so the design declines it outright, the same guard persona reviewers would later praise on Developer Profile.
Reframe rather than invent."Value interpretation" became spend-by-deliverable-state instead of a made-up value score.
Nothing silently dropped. A coverage table maps every client pain point and one-pager promise to MVP, Phase 2, or a flagged decision.
Numbers are only impactful with context. See something → investigate it.
Self-review by annotation
No formal audit existed yet, but the wireframe self-reviews, with the highest density of self-flagged weak takeaways of any stream: three, each with a written fix, plus seven open questions listed explicitly, including the budget-source decision blocking two panels' denominators.
55
Source tags
8
Takeaway strips
3
Self-flagged weak
7
Open questions