AI sales-intelligence platform
LeadForge AI
Find the right businesses. Understand them. Start the right conversation.
- Role
- Solo · design, build and ship
- Focus
- AI · Web
- 198
- automated checks
- 6
- scoring dimensions
- 46
- design documents
- $0
- to run end to end
Overview
Researches local businesses, explains why each one is worth contacting — with citable evidence for every claim — and drafts the message you would actually send. A claim the model can’t trace back to a stored observation is dropped before it ever reaches you.
What it does
- Ranked shortlist with a reason for each lead — every claim linked to stored evidence
- Observed, Inferred and Unknown labelled explicitly; missing data is never invented
- Deterministic scores across ICP fit, opportunity, contactability, maturity, confidence and urgency
- Messages validated by a deterministic rule engine and an adversarial model check
- Manual channel with one-click wa.me / mailto — nothing sends without approval
Engineering decisions
A claim without a citation doesn’t ship
The model is asked to explain numbers computed in code; any numbers it returns are discarded.
The worker boots the API’s own module graph
Verification, scoring and outreach run from HTTP and the queue through one implementation, so a lead can never score differently depending on the path.
SSRF protection on every redirect hop
Private, loopback and cloud-metadata ranges are refused before each hop, with byte and time caps.
Rate limits reschedule rather than retry
A daily quota no longer burns five retries in thirty seconds and dead-letters while hours of quota remain.
Stack
- Frontend
- Next.js 15React 19TailwindRadixTanStack Query
- Backend
- NestJS 11Prisma 6Zodargon2id
- Data
- PostgreSQL 16Redis 7BullMQ 5
- AI
- OpenRouterStrict JSON schemas