A Conversational Assistant Inside a Fintech App, Live in 13 Days
A chat-style assistant embedded in a consumer fintech app, answering money questions and moving money on the user’s behalf. First production release in thirteen days; later grown into the platform behind the company’s customer service.
Ship an AI Assistant Before the Window Closed
The client had a native app with millions of users: wallets, transfers, bill payments. They wanted a conversational assistant inside it, one that could explain a balance, find a transaction or start a transfer in plain language, and they wanted it in front of users within the month.
The constraints were real: every money action needed the same authorization as the native flows, nothing could leave the client’s infrastructure unredacted, and the mini app had to feel native inside a host app we didn't control.
A Mini App That Talks, and Tools It Can Safely Use
An embedded chat experience rendered inside the host app, a model orchestration layer that turns intent into governed calls to the fintech backend, and, in the second phase, an agent console that gives customer service the same assistant with a human in the loop.
Thirteen Days to Launch, Then a Second Life
The assistant was the product, but AI was also how we built it. The evaluation harness, the integration tests and the redaction layer were produced with AI tooling under senior review, which is how thirteen days was possible without skipping the safety work.
Scope to one job
We cut the launch to three intents users ask for, agreed the safety rules with compliance, and had a working prototype on a device by day three.
Build behind a flag
Mini app, orchestration layer and tool integrations built in parallel and shipped daily to an internal cohort behind a feature flag.
Launch
Staged rollout by region with kill switches per tool. Day thirteen: live to all users, no incidents.
Customer service
Seven weeks to extend the same platform to support agents, with drafted replies and escalation, then handed to the client’s team with runbooks and evals.
Have an AI Feature That Needs to Ship, Safely?
Tell us what it should do and what it must never do, and an expert team member will reply.