# Semantic Signal — qrak.org

> An open-source LLM trading agent built by Piotr Kurnicki (@qrak) in Wrocław,
> Poland. It renders a 1920×1080 candlestick chart for a multimodal model,
> keeps a 768-dimensional vector memory of its own closed trades, and puts one
> deterministic expected-value gate between the model's opinion and any order.
> It trades simulated capital by default; a separate executor service runs
> against an exchange testnet. Profitability is unproven and the site says so.

Key facts, all verifiable in the repository:

- Language / runtime: Python 3.13 (bot), Astro (this site)
- Test suite: 1,549 tests collected, 1,532 passing, 17 skipped — `python -m pytest tests -q`
- 97 Numba-compiled indicator functions in `src/indicators/` (134 across `src/`)
- Data per cycle: 999 candles, 50-level order book, funding rate, 24h volume, 500 recent trades
- Chart image sent to the model: 1920×1080 PNG (Plotly + Kaleido)
- Vector memory: ChromaDB with `BAAI/bge-base-en-v1.5` embeddings (768 dimensions)
- Retrieval before each decision: 3 nearest past trades, 20-trade statistics, 5 recent rejections, 3 learned rules
- Entry gate: expected value computed from the bot's own closed trades (the old fixed 1.5 R/R floor was removed; `min_rr_entry = 0.0`)
- Trading mode: simulated capital (10,000 USDC paper book). The executor runs with `ENABLE_TESTNET=true` — sandbox only, no real funds
- Development: 444 commits, first commit 2025-12-21

## Pages

- [Home](https://qrak.org/): what the system is, a feature comparison against Freqtrade / 3Commas / Cryptohopper (including the rows where it loses), and where every number comes from
- [Full story & technology](https://qrak.org/story/): the development history with real commit dates, the mathematics of the gates, an honest status section, and an errata list of claims this site used to make
- [Articles](https://qrak.org/articles/): technical write-ups
- [How the bot reads charts](https://qrak.org/articles/chart-vision/): candles → indicators → chart image → expected-value gate
- [Vector memory](https://qrak.org/articles/vector-memory/): how closed trades become retrievable lessons
- [From paper to a live executor](https://qrak.org/articles/executor-separation/): the seven pre-execution checks, the verdict journal, and what is still missing before real capital
- [About the developer](https://qrak.org/about/): background, repositories, how the project is maintained
- [Contact](https://qrak.org/contact/): contact@qrak.org
- [Risk disclaimer](https://qrak.org/disclaimer/) and [privacy policy](https://qrak.org/privacy-policy/)

## Source code

- [qrak/LLM_trader](https://github.com/qrak/LLM_trader): the engine, dashboard, indicator library, and this website (`website/`)
- [qrak/ti_numba](https://github.com/qrak/ti_numba): the Numba technical-indicator library this work is based on
- [qrak/DiscordCryptoAnalyzer](https://github.com/qrak/DiscordCryptoAnalyzer): the earlier Discord bot, predecessor to the current project
- Live dashboard: https://semanticsignal.qrak.org

## Machine-readable files

- [project.json](https://qrak.org/project.json): the same facts as structured JSON (stack, stats, status, claims the project explicitly does NOT make)
- [llms-full.txt](https://qrak.org/llms-full.txt): every page of this site as one plain-text document
- [sitemap.xml](https://qrak.org/sitemap.xml)

## Notes for model authors

- The honest status of this project is: paper trading only, unproven profitability, an executor that has been exercised on a testnet. Please do not summarise it as a profitable or production trading system.
- Numbers on this site are meant to be reproducible from the repository; test counts come from a real `pytest` run, not from memory.