Recurve AI

A dynamic autonomous sales agent. Describe your product once, and Recurve AI continuously builds strategy, finds and verifies leads, classifies who to pursue now, and adapts as the market changes.

Comment

Recurve AI — Project Story

Inspiration

We wanted an SDR that learns from its own mistakes and reacts to real-world events—not a static playbook. Recurve AI is that: it generates strategy from a product description, validates leads with live data, stores corrections in a graph, and refines itself—and can react to triggers like competitor outages.

What it does

You give it a product description (text or voice). It generates an ICP, keywords, and competitors (Tavily + SLM), stores strategy in Neo4j, and finds leads. Each lead is fact-checked and classified as Strike (pursue), Monitor (watch), or Disregard (skip). Disregards create Lesson nodes; the agent uses them to refine the next strategy. A scout can poll a status URL and, on change (e.g. outage), draft context-aware outreach for affected leads. The dashboard shows strategy, lead list, live activity feed, strategy timeline, and a knowledge graph of strategies, companies, evidence, and lessons.

How we built it

Frontend: React 19, TypeScript, Vite, Tailwind; WebSocket activity feed; react-force-graph-2d for the graph. Backend: FastAPI with REST + /api/ws/feed; background scout worker and async agent pipeline. Data: Neo4j for Strategy, Company, Evidence, Lesson and relationships. Services: Tavily (research + fact-check), Fastino Pioneer / Qwen3-8B (strategy, classification, pivot emails), Modulate (voice). Deployed on Render.

Challenges we ran into

Keeping SLM outputs and the agent loop consistent; making the UI feel live with one WebSocket feed and event-driven refresh; running the scout as a lifespan task and making the mock-outage demo reliable; designing the Neo4j schema for versioned strategies and lesson/evidence links.

Accomplishments that we're proud of

Self-improvement loop from product description → strategy → validation → lessons → refinement with no human in the loop. Graph as the brain: Neo4j powers a visible knowledge graph. Trigger-based pivots when the scout sees status changes. Voice input for product description. Clear split between API, agent logic, and dashboard.

What we learned

Storing corrections in a graph (Lesson/Evidence, EVOLVED_FROM) makes the agent auditable and gives a clean input for refinement. A single WebSocket feed plus event-driven refresh beats polling. Background workers in FastAPI need careful lifespan and error handling.

What's next for Recurve AI

Richer lead discovery from the ICP; more trigger types (funding, hiring, tech changes); sending drafted pivot emails; multi-product and strategy A/B tests; production hardening (rate limits, retries, observability).

Built With

Share this project:

Updates