A trading workstation on a live Argentine broker API: a rebalance engine turns analyst signals into orders, and nothing reaches the broker until a human confirms it. The analyst backend is a fork of the open-source ai-hedge-fund project.
Layers built
Sole author of the desk: the Next.js frontend, the broker integration and the rebalance engine. Contributor to the forked Python backend.
Python, FastAPI, LangGraph, Anthropic, Next.js, React, TypeScript and Zustand.
Two repositories. The desk is my own work. Its analyst backend is a fork of the open-source ai-hedge-fund project by virattt: the 12-agent analyst panel and the LangGraph orchestration are upstream, not my design. My 72 commits there adapt it to the Argentine CEDEAR market and add the rebalancer, the backtesting layer and liquidity ranking.
A trading workstation wired into a real Argentine broker API (InvertirOnline). The desk reads the analyst backend's signals, works out a rebalance, and then stops. It lists every order it wants to place and waits. Only a confirm click reaches the broker, and in demo mode that same call returns a fake ticket, so the public demo executes nothing.
The analyst backend is a fork of the open-source ai-hedge-fund project. Its 12-agent panel and the LangGraph orchestration are upstream work, not mine. My 72 commits there adapt it to the Argentine CEDEAR market and add a rebalancer CLI with consensus ranking, a backtesting layer that sits behind the same broker interface, a point-in-time liquidity ranker and commission-inclusive cost basis.
The frontend reframes the whole thing as a nostalgic desktop OS. Portfolio, live market data, quick-trade, auto-trader and backtesting are each their own draggable window, built on a hand-rolled Win98-style window manager.