The challenge
Flexible conversation needs a dependable path from intent to data, interface, and action.
Portfolio · AI product design
I explored how to give a finance agent a controlled way to interpret intent, choose a workflow, and return an interface people can inspect.
The challenge
Flexible conversation needs a dependable path from intent to data, interface, and action.
My approach
Let the Agent choose and compose supported components. Define their data, states, and confirmation rules.
The exploration
A working implementation of intent-led responses, component contracts, and reviewable actions.
01 / My role
Mono is my independent exploration of conversational personal finance. I owned the product flows, UI system, component contracts, and implementation, from interpreting a request to confirming a record.
Define how intent, missing information, and user decisions shape the workflow.
Specify supported components, required data, and user actions.
Connect visual rules and reusable cards to a working renderer.
02 / Design direction
I compared three ways to give a model a role in the UI. My design problem was consistency: each request may need a different answer, while amounts, time ranges, and save actions need a stable structure.
Predictable presentation, with composition limited to predefined screens.
Flexible answers within a designed vocabulary of data and interaction contracts.
More presentation freedom, with a larger validation and execution surface.
I chose declarative composition. Information architecture connects the task to a component; progressive disclosure lets the answer expose the details needed for that task. This direction came from design and technical exploration.
03 / Agent workflow
I structured the Agent around a LangChain intent router and task-specific chains. Product code dispatches the request, retrieves or calculates the required data, and prepares a supported UI response.
Intent routing. The request and available context enter LangChain, which returns a structured intent for product code to dispatch.
When required information is missing, the workflow asks a follow-up and carries the request forward. The model does not have to finish the task in a single response.
04 / UI contract
I defined the handoff as intent, component, data, and feedback. Schemas constrain the output; the component vocabulary defines what can be rendered and which interactions it supports.
Preserve the request’s meaning and produce structured intent and task output.
Provide records, context, and calculations to the response.
Map supported output to registry components or dedicated business views.
The interface vocabulary includes what a component can do, not just how it looks.
I paired visual tokens with component names, props, data requirements, and action rules. These make the design system usable as an implementation contract.
05 / Human handoff
I placed an explicit review between Agent output and a transaction write. The person can inspect and edit the proposed values; product code submits the change only after confirmation.
06 / Exploration outcomes
Building Mono let me test the design idea in working software: a request can lead to different interfaces while each response follows product-defined data and action rules.
Intent routes to an executable workflow with a supported response contract.
Schemas, component props, and rendering rules turn the vocabulary into working UI.
The Agent composes supported responses; product code controls rendering and saving.
The limit is expressiveness: a supported vocabulary works only for the tasks it can represent. New intents need another component, a clarification, or a fallback.
07 / Reflection
Mono made that responsibility concrete: define the vocabulary, connect it to product data, and decide where a person reviews an action. The next question is how that vocabulary should grow: when to add a component, when to ask for clarification, and how to respond when the current set cannot express the request.
The live app requires sign-in. The recording and workflow diagrams above can be viewed without an account.