Fintech Banking GTM Engineering
A signal-based selling engine for a business banking platform: it detects a funding round or a competitor complaint, scores the company on fit, timing, access, and intent, and drafts the outreach for a person to approve.
Overview
Business banking is won at a moment: the week a startup closes a round, or the day a founder complains in public about their current bank. This engine watches for both. Funding announcements arrive over RSS and a manual live trigger, competitor-displacement complaints are captured the same way, and Claude scores every company on four axes, 0 to 25 each: Fit (is this who the platform serves), Timing (are they choosing a bank right now), Access (can you reach the person who decides), and Intent (are they actively unhappy or actively shopping). Anything at 60 or above gets two drafts, one for the founder and one for the finance lead, each opening on the specific trigger instead of a generic congratulations, and nothing sends until a person approves it. Signals to call list is the verified path and runs live end to end. The outbound search, sequence builder, monitor library, and agent studio around that loop are built, and the pieces not yet wired to a provider say so where they appear. Every company in the demo is sample or public data, and the account values are labeled estimates with their assumptions shown next to the number.
What it involved
- Two signal types, one loop: funding rounds and competitor-displacement complaints are captured, scored, and drafted in under a minute from headline to email.
- Four-axis ICP scoring (Fit, Timing, Access, Intent, 0 to 25 each) with a rubric matched to the signal type, the sum enforced in code, and a plain-English rationale on every company.
- Persona-specific drafting: a founder email and a finance-lead email per company, each opening on the trigger, under 120 words, with the hard constraints checked after generation.
- Human review before any handoff: a person approves or rejects every draft, and nothing auto-sends.
- Apollo-connected outbound search that turns a plain-English ICP into filters and pushes the candidates into the same call list.
- Extend it without an engineer: an agent studio and a monitor builder where a rep describes what they want in plain English and the system writes the config.
- Account-value estimates priced off deposits and card spend, with the modeling assumptions cited next to every number.
Stack
- Next.js 16
- React 19
- TypeScript
- Tailwind v4
- Supabase Postgres
- Claude (scoring, drafting, assistant)
- Apollo
- Vercel cron