Agentic AI · Interaction system · ProNovos 2.0

Nova

An autonomous presence people can trust, and know when to ignore.
Role
Design lead + front-end, 0 → 1
Scope
Interaction system · doctrine · vocabulary
Timeline
2025 → ongoing
In production at ProNovos
Info

Every product grew a chat bubble. Nova is the opposite bet: an AI designed into the body of a construction-finance platform, including its layout, its language, and its safety model, so thoroughly that "where does the AI live" stops being a question. This is the case study of never feeling bolted on.

01 · The bolt-on problem

LLM chatbot in a trench coat ≠ AI-first product

When we started talking about building the new product, the main thing the Director of Product and I agreed on was this: it would never, ever, ever be a 'chatbot.' You know, the kind that just sits there waiting to be called on, gives an answer, and retires.

The original ProNovos product was built with an action-first mentality: if something wasn't actionable, we didn't need it. That belief is even more important when building an AI product. AI significantly expanded the range of actions the product could execute, and users expect to do far less manual work than they used to with traditional apps.

02 · Anchored, not floating

Signals live where the data lives.

A bolted-on AI chatbot tells users to go do something somewhere. Nova goes to the right place so that the recommended action is just a click away, or it performs an action for the user and lets them edit or undo.

Every page element that matters carries an anchor, a signal docks its peek to the anchor, highlights related cells, and renders the proposed value as a ghost edit inside the cell it would change.

No information is displayed without an action. A user may choose to view a dashboard the "traditional" way, as Nova never shows up like without a reason like an annoying pop-up. But it is there where user selects an action or a KPI, suggesting an action or providing more context.

03 · Context

Pointing replaces describing.

The second bolt-on failure, the AI can't see the screen, dies here. In ProNovos, Alt+click anything and it becomes a reference chip in Nova's composer. Not a screenshot guess: components register ground-truth metadata with a probe registry: kind, name, metric id, current value, project scope, and the underlying series. The user asks "why is this like that?" without spelling out what this is.

The same capture path powers three entrances: Alt+click on any element, a select-text → "Ask Nova" pill on any highlighted phrase, and a persistent "?" beside every page heading to explain the view.

04 · The body

The page makes room. The panel never covers.

Nova's panel drags anywhere and snaps magnetically to an edge. When it docks, it doesn't overlay. The app root pads itself via CSS variables and the content genuinely reflows.

07 · The language

Users never meet the word "orb."

Specs sayUsers see
orb · peek · seam
Nova
signal → flag → brief → alert
Nova noticed
ghost edit · receipt
Ask Nova
loudness ladder

Integration is also a writing problem. I run a two-tier vocabulary: internal spec language that lets tickets stop saying "chatbot" and use signal, flag, brief, alert, orb, peek, seam, and ladder; and a user-facing layer where all of that machinery collapses into one word: Nova. The orb is never product copy. To users, Nova simply is there, the way a colleague is there.

08 · Where it goes

From feature to platform.

Nova Canvas lets you prompt a bespoke dashboard into existence, then the chat docks where the configure panel was and you talk it into shape. Safe because we secure the cage, not the code: the AI authors presentation only, never data access; generated React runs in a no-network sandbox against a tenant-scoped payload the server resolved.

09 · more work