Freqtrade, 3Commas, Cryptohopper — None of Them See What This Bot Sees
An autonomous crypto trading agent with multimodal 1080p chart vision, 768D vector trade memory, self-written rules, and deterministic expected value gates.
The Technical Scoreboard
| Capability | Freqtrade | 3Commas | Cryptohopper | Semantic Signal (Mine) |
|---|---|---|---|---|
| Chart Vision (reads image) | ✘ | ✘ | ✘ | ✔ (Gemini Flash) |
| Semantic Trade Memory | ✘ | ✘ | ✘ | ✔ (ChromaDB 768D) |
| Bull/Bear Self-Debate | ✘ | ✘ | ✘ | ✔ (Single-Prompt) |
| EV + Falsification Gates | ✘ | ✘ | ✘ | ✔ (Deterministic Math) |
| Hardened CCXT Executor | ✔ | ✔ | ✔ | ✔ (llm_trader_executor) |
| Self-Maintaining Codebase | ✘ | ✘ | ✘ | ✔ (8-Agent System) |
| Zero-Cost News & Sentiment | ✘ | ✘ | ✘ | ✔ (RSS + Reddit RAG) |
Multimodal Chart Vision
Renders 1080p Plotly candlestick charts directly to Gemini Flash. The model reads visual geometry (support wicks, volume spikes, wedge breakouts) rather than flat indicator floats.
768D Vector Memory
Every closed trade is embedded into ChromaDB (BAAI/bge-base-en-v1.5). Queries top-5 similar past trade setups and applies the Surprise Ratio metric to filter out market noise.
EV & Falsification Gates
Calculates Kelly Criterion position sizes based on historical win rates and forces the LLM to write an explicit price invalidation trigger before any signal is accepted.
Decoupled Execution
Separates reasoning from live execution. The main engine outputs atomic JSON decisions to llm_trader_executor, protecting exchange-side OCO stop-losses from API latency.
Built in a Wrocław Apartment After Warehouse Shifts
In December 2025, I was working 8-hour warehouse shifts in Poland. Tired of subscription "AI bots" that were just simple moving average crossovers, I spent 7 months building a cognitive trading engine with 1,270+ automated tests.
Read the Full 7-Month Development History →From the Blog
📈 How the Bot Reads Charts
999 candles, 50+ Numba indicators, a 1080p chart image — and an LLM that has to defend its call.
🧠 Vector Memory
How 768D embeddings turn closed trades into lessons the bot retrieves before every decision.
🛡️ Paper → Real: The Executor
Why reasoning and order execution are separate services, and how the system earns real money.
⚠️ Risk disclaimer: Semantic Signal is an experimental, open-source software project — not financial advice, not a money machine. It currently trades simulated capital by default, and trading cryptocurrencies carries substantial risk of loss. Read the full disclaimer.