Esmeralda

Context

Trading futures by hand has two problems: discipline and context. Discipline can be solved with rules, and rules can be programmed. Context is harder: a technical signal that looks perfect can be a trap if the portfolio is already overexposed, or if a piece of news just changed the picture. Esmeralda is my attempt to attack both in the same system.

What I built

A full-stack platform that runs continuously rather than in batches or on request:

• Continuous Binance Futures ingestion. The system consumes market data without interruption and normalizes it so strategies always work from the same baseline.
• Configurable strategies evaluated across different risk profiles and timeframes, so the same signal is judged differently depending on how much risk is acceptable.
• An LLM-assisted decision layer that evaluates each signal with full context: current position, portfolio state, accumulated risk and financial news.
• A PWA with push notifications and Discord alerts, so a decision arrives immediately without having to watch a screen.
• Continuous supervised live testing, which is the only honest way to learn whether a strategy holds up outside of a backtest.

Technical decisions

• Separating the signal from the decision. Strategies produce deterministic, reproducible signals; the LLM does not generate the signal, it judges it with context. That keeps the system auditable: you can always see what triggered a signal, separately from why it was accepted or discarded.
• Risk is a first-class parameter, not a filter at the end. A risk profile changes which strategies apply and how a position is sized.
• Python for ingestion and evaluation, React for the interface, PostgreSQL as the record of signals, decisions and outcomes: without persistent history there is no way to measure whether the system is improving.
• Supervision before full automation. It runs live, but with a person in the loop; the goal is to validate the judgment before delegating anything to it.

Result

A continuously operating system that combines real-time market ingestion, strategy evaluation by risk profile, and a decision layer with real financial context, delivering alerts through a PWA and Discord.

It is in development and under supervised live testing. It is not financial advice and not a commercial product: it is a research and decision-support system.