ProspectAI
A 6-agent CrewAI pipeline for stock research
Hi IH I’m Moises, an AI engineer in Barcelona. I want to share a side project I built to learn agentic AI.
I built ProspectAI a multi-agentic system with CrewAI. It’s an open-source pipeline of 6 CrewAI agents (market, technical and fundamental analysts → a draft strategist → a critic → a final strategist) that researches a stock portfolio end to end.
Three decisions did most of the work, and they’re the part I think is reusable beyond finance:
1. Keep the LLM away from arithmetic. The model only decides which signals matter; every number is computed by deterministic Python tools. The moment an LLM does math, errors compound silently downstream.
2. Typed contracts between agents beat freeform text. Passing Pydantic-validated objects instead of prose killed a whole class of bugs where one agent reinterpreted another’s narrative.
3. A dedicated adversarial critic catches what a single pass won’t. One agent’s only job is to attack the draft and issue structured fixes.
It’s a learning/demo project on public data — explicitly not investment advice. MIT-licensed, and the live demo streams the whole run token-by-token over SSE.
Demo: prospect-ai.moisesprat.dev
Code: github.com/moisesprat/ProspectAI
Happy to share with the community and to get feedback. Tokens are at my cost, so any donation would be appreciated.
1 Comment
The typed contracts point really stood out.
It feels like a lot of multi-agent failures get blamed on the models themselves, when they're really communication failures between agents.
Treating those handoffs more like software interfaces than conversations seems like a much more reliable direction.