27 agent tools
Tool use covered the requested domain actions and information paths. Tool availability was tied to the operating mode.
Delivered artifact · three-day take-home
A dual-mode financial AI agent with tools, fallback, and audit boundaries.
This record demonstrates a completed technical build under a constrained evaluation. It is not presented as a production deployment, compliance certification, or measured client ROI case.
The brief translated into a system
One mobile application, two operating contexts.
The implementation used React Native and Expo. A customer-facing mode and an internal-team mode exposed different tool surfaces instead of giving every user the same agent and permissions.
Provider fallback
Gemini 2.5 Flash was configured as primary with Llama 3.1 on Groq as an environment-controlled fallback.
Privacy-aware design
The build included anonymous operation and a hashed audit-chain design against the stated privacy and lending constraints.
Implementation evidence
- A working mobile codebase was delivered
- Customer and internal modes were implemented
- Twenty-seven tools were wired into the agent
- Provider fallback was designed
- The reviewer response was followed by collaborator invitations
Production evidence not available here
- No public operating client deployment
- No independent compliance certification
- No measured user adoption or financial outcome
- No public long-duration reliability record
- No claim that take-home behavior equals production readiness
What this demonstrates
Fast implementation still needs explicit boundaries.
The strongest signal is not feature count. It is the ability to translate a regulated-domain brief into user modes, tool permissions, fallback behavior, audit concepts, and a working artifact under time pressure.
For an actual production build
Add threat modeling, evaluation data, access tests, monitoring, and domain review.
A real deployment would begin by converting policy and operating risk into acceptance tests and named approvals.