Horizon Desk
Built with a three-person team at SteelHacks XIII, Horizon Desk uses synthetic banking data and AI-assisted outreach to identify customer opportunities, with deterministic eligibility checks and human review for credit decisions.
- Role
- Frontend + supporting backend
- Team
- 3 developers
- Event
- SteelHacks XIII
- Year
- 2026
- Status
- Completed prototype
- Stack
- Next.js
- TypeScript
- Claude 3.5 Sonnet
- ElevenLabs
- Python
- Vitest
Problem
Retail banks can hold large balances in low- or zero-interest checking accounts, and branch bankers have limited ways to spot which customers may have surplus liquidity and reach them in a compliant, personal way.
The team prototype
Our team built Horizon Desk for SteelHacks XIII. The prototype analyzes synthetic retail-banking customer data, identifies potential surplus liquidity using deterministic financial eligibility checks, and drafts personalized AI-assisted outreach. It also simulates customer responses and outreach cooldowns, and routes credit opportunities to human bankers rather than automatically approving loans.
My contribution
I helped originate the project concept, worked primarily on the frontend, and contributed to some of the backend development.
Team project
Team
- Frank Ncube
- Gamuchirai Mubayiwa
- Sumon Mondal
Horizon Desk was built by a team of three. The prototype’s functionality reflects shared team work rather than any one contributor.

Built at SteelHacks XIII
University of Pittsburgh · 2026

Architecture
A simplified view of the prototype’s main workflow. Customer records, responses, and outcomes are synthetic or simulated.
Engineering decisions
01
Deterministic eligibility before AI-assisted outreach
Potential surplus liquidity is identified with deterministic financial eligibility checks rather than a language model, so who is contacted follows explicit criteria. AI-assisted drafting comes afterwards.
02
Human review for credit opportunities
Credit opportunities are routed to human bankers instead of being approved automatically, which keeps lending decisions with people.
03
Synthetic data and modeled outcomes stay labeled
The prototype runs on synthetic customer data and simulated responses, and it is presented as modeled behavior, not as real customer activity or results.
Scope and limits
- All customer records and financial projections in the prototype are synthetic or simulated.
- Integration with core banking systems was a planned next step and was not implemented.