AI agents combined crypto market signals with token conversations

A crypto trading firm making real-money decisions on noisy market data. Crypto never closes, so the feeds never stop and there is no overnight window to catch a bad assumption before the firm trades on it. I came in from high-budget, high-volume advertising data work, aggregating many sources and running attribution, and trading data was more intensive than all of it.

Market data shows what price did. Reading what's happening around a token also takes the crowd: the top influencers and the projects' own accounts on Twitter, everyone talking about the project on YouTube, and the project's own Discord. Reading all of that and acting on it within a couple of minutes takes a team of analysts, and analysts can't work 24/7.

Traditional machine learning handled the numbers: time-series market data ran through technical indicators, the indicators rolled up into compound scores, and the scores fed regime detection. Scrapers pulled the Twitter, YouTube and Discord feeds. Then all of it was merged, the way a customer data platform builds one profile per customer, into one profile per token, and handed to a council of agents. Each agent had its own personality, background and tools, read a different slice and wrote its report in its own format. The reports went up to other agents, which checked the claims, ran their own research when something looked off, and made continuous decisions that adjusted the regime detection. Every insight was saved.

• Python
• SQL
• Supabase
• LangGraph
• Technical and custom indicators
• Machine learning for regime detection
• Twitter, YouTube and Discord scrapers
• Multimodal LLM agents

The firm got a read of the market and the crowd that kept running when people would have gone home: quantitative signals and the conversation around each token in one place, and a council that balanced the models rather than taking them on trust. Because every insight was saved, the team could check the models and the agents against what the market did.

Doing the same reading with people means hiring a team of analysts who can't work 24/7 or respond within a couple of minutes. The council did it around the clock, and its saved insights turned its own judgment into a time series that can be checked for accuracy.

Give each job to the tool that is good at it. Traditional machine learning is the right engine for indicators and regime detection; agents are the right engine for reading text and indicators together and pulling many reports into one balanced view. The hard part, and the fun part, was merging truly multimodal data. My data-engineering background made it possible, and so did multimodal agents that read text and quantitative indicators side by side.