Everything here is embedded live rather than exported to an image. Trace a node through its lineage, or open the console prototype and walk a recommendation through to approval. Each one carries the reasoning that shaped it.
2 pieces
Databricks Cost Intelligence
A read-only control loop over an existing Databricks estate. It collects telemetry, builds a workload digital twin, generates constrained optimisation candidates, simulates their cost and performance, then explains the recommendation and routes it through human approval before anything changes.
Design notes
Collection is strictly read-only. The system never holds write credentials against the production estate, so the worst case failure of the analysis layer is a bad suggestion rather than a bad mutation.
Candidates are generated inside explicit constraints rather than by free-form search. Bounding the space up front is what keeps the simulation step cheap enough to run continuously.
Every recommendation carries its own explanation and cost delta. An unexplained number does not get approved, so the explanation is part of the artifact rather than a downstream report.
The approval gate is deliberately human. Actions are applied through a controlled path with an audit trail, which keeps the loop governable as it widens.
Stack
- Databricks
- Unity Catalog
- System Tables
- LLM
- FinOps
Cost Intelligence Console
The approval surface from that architecture, built as a working prototype. An operator scans the ranked candidates, reads the evidence behind one, checks the simulated cost and risk, then approves it into a canary rollout. Every control is live, so the interaction model can be judged before any of it is built for real.
Design notes
The screen answers one question in order: what should change, what is that worth, and what could go wrong. Ranked list on the left, the case for a single candidate on the right, approval pinned to the bottom where the decision ends.
Colour carries meaning rather than decoration. Green is reserved for money saved, amber for anything a reviewer has to weigh, and the single red accent marks the active row. Everything else is grey so the numbers stay the loudest thing on screen.
Money and identifiers are set in a monospace face with tabular figures, so a column of savings stays comparable at a glance and digits never shift as values update.
Evidence, simulation and risk sit behind tabs instead of one long column. A reviewer looking for the risk case should not have to scroll past the evidence to reach it.
Confidence is encoded twice, as a bar and as a number, because a bar alone is hard to compare across rows and a number alone is easy to skim past.
Nothing on the screen can mutate the estate. The read-only badge, the approval lock and the canary wording all restate the same guarantee the architecture makes.
Stack
- Databricks Apps
- Design System
- Dark UI
- Data Density