The problem
Crypto Twitter is downstream of Telegram. The early conviction on a token, the hint of a partnership, the panic about a depeg — it shows up in Telegram groups hours before it’s a tweet thread. Useful, but only if you’re willing to watch dozens of channels at once. Nobody is.
What we’re building
A Python service that joins configured Telegram groups via the Bot API, ingests every message, and runs each through a sentiment + topic scorer. Per-token rolling sentiment is correlated against on-chain price movement on a short lag — the system surfaces channels that have predictive value, not just noise.
The AI angle
Sentiment is the easy part. The hard part is signal extraction: distinguishing “genuine early conviction” from “coordinated shilling”, separating sarcasm from earnest panic, attributing claims to credible authors. Claude scores each message with structured output (sentiment, confidence, topic, claim type) and the aggregator does the rest.
How it’ll be used
- Crypto traders who want a single dashboard instead of fifty pinned channels.
- Analysts and researchers studying sentiment-vs-price relationships at scale.
- Internal R&D — we use it as a sandbox for evaluating LLM-driven sentiment systems.
Where we are
Ingest, scoring, and the correlation pipeline are working. We’re running it against historical data to validate which channels have measurable signal before opening it up beyond internal use.