AI Dashboard

AI capability is easy to demonstrate and hard to deliver. A demo runs in ideal conditions on its own page; a product has to run inside a host application that was not designed with it in mind.

This one is not a chat box either — it is a decision surface. It takes a client's financial picture and answers a specific question: which tax strategies apply to this person, and what is each one actually worth?

That means seventeen distinct strategies, from an S-corp election to the Augusta Loophole to hiring your child, each with its own eligibility criteria, its own required inputs, and its own savings calculation — all resolving into one ranked view with an engagement plan and concrete next actions at the end.

A host page is an unpredictable environment. Global CSS leaks in both directions, script load order is not ours to control, and the surrounding DOM can be anything at all. Weight matters too: an embed that ships a heavy bundle degrades the host's performance, and the host owns that experience.

The harder constraint is that the tax logic will not hold still. Rules change with legislation, thresholds move each year, and the strategy list only grows. Encoding that logic in the frontend would make every rule change a frontend release — the exact coupling that turns a product into a maintenance burden.

The strategies also have almost nothing in common structurally. Some need a single number, some need a full business profile, some depend on the answer to another strategy. Building seventeen bespoke UIs was never going to scale to the eighteenth.

Push the strategy engine out of the interface entirely and orchestrate it as n8n flows — one pipeline per strategy, each independently editable. The frontend renders results; it does not know tax law. A rule change becomes a flow change rather than a release.

Make every strategy return the same shape across that boundary: eligibility, required inputs, projected saving, and the next action it implies. Seventeen strategies then render through one component set, and the eighteenth needs no frontend work at all.

Ship the dashboard as Web Components, so style and DOM isolation are enforced by the platform rather than by convention, behind a small documented element plus a typed configuration contract that is identical for every host.

Design the workflows AI-first — streaming responses, in-flight states, and partial results treated as normal operating conditions rather than edge cases bolted on afterwards. A strategy engine that fans out across flows is never uniformly 'done', so the UI is built to show progress honestly.

Hold a hard line on rendering performance, because an embedded surface is judged as part of the host application, not as a separate product.

The AI capability became something a platform can actually adopt: embeddable, consistent, and light enough that the host does not pay for it in perceived performance.

Seventeen strategies ship behind one interface, and the strategy catalog can grow without touching the frontend — the orchestration layer absorbs the change.

The output is the part that matters to the client: not a list of suggestions, but a ranked set of opportunities with projected savings attached and an engagement plan telling them what to do next.