I am a weak trader. I am not a weak automator. So I stopped pretending to have iron nerves and built an AI system that executes rules more calmly than I do and reviews mistakes after every trade.
What went wrong with manual trading
For a year and a half I studied crypto: futures, spot, arbitrage and bots. I understood the mechanics. But I did not become a good manual trader.
The problem was not only knowledge. The problem was behavior. When a drawdown starts, I get nervous, close too early, break my own rules and then watch the setup work without me. Lost deposits, nerves and time.
Crypto amplifies this problem: it runs 24/7, is volatile, noisy and emotional. The SEC warns that investments in crypto asset securities can be exceptionally volatile and speculative.
“Investments in crypto asset securities can be exceptionally volatile and speculative.” — SEC Investor.gov
Simple conclusion: when the environment is volatile and the human is emotional, manual execution becomes the weak link.
Why I chose automation
I combined what interests me with what I actually know how to do: AI + crypto + automation.
The idea is not that AI “predicts the market.” That is a dangerous fantasy. The idea is more modest: a system should execute predefined logic without panic, fatigue or impulse decisions.
What an algorithm does better than me:
- it does not get tired of watching the market;
- it does not get angry after a stop;
- it does not revenge-trade;
- it does not change the strategy out of fear;
- it logs data for later review.
How the first version works
The pet project flow is:
- Market signals arrive: open interest, liquidations and volume.
- Algorithms evaluate the situation and filter noise.
- An LLM analyzes context and decides whether to enter or skip.
- If there is an entry, the system defines stop, take profit and position size.
- After the trade, a separate review extracts lessons for future setups.
The first stack: DeepSeek R1, Windmill, Bybit and Telegram. At the time of the original post, the monthly win rate was around 53%. That is not proof of profitability and not a trading signal. It is only one metric of an experimental system, and it must be viewed together with risk/reward, drawdown, fees and sample size.
The main part is Lesson Learning
The most interesting part for me is not the entry. It is the review after the trade. If a system only repeats human mistakes faster, that is bad automation. So after every trade, a separate model analyzes what happened: which signal worked, which threshold was too loose, where the setup was weak, and where the market changed.
NIST’s framing is useful here: AI systems should be reliable, transparent and governed, especially when they influence risky decisions.
“Characteristics of trustworthy AI systems include: valid and reliable, safe, secure and resilient, accountable and transparent...” — NIST AI RMF
Plain English: if AI participates in trading, you need to understand the input data, logic, limits and responsibility. Otherwise it is not a system; it is a casino with an API.
The business lesson
Crypto is not the main topic here. The main topic is automating a weak spot.
In many business processes, a person knows the rule but executes it badly: forgets, rushes, reacts emotionally or fails to log outcomes. AI and algorithms are useful where discipline, repeatability and feedback loops matter.
Honest limits
AI trading does not remove risk. A model can be wrong, data can lag, an exchange can fail, a strategy can decay, and a nice backtest can break in the real market.
A serious version of this project should start not with “how much will it earn?” but with “what is the maximum risk, what are the stops, what is logged, what sample size is enough, and when do we shut the system down?” For me, this is a pet project and an automation lab, not a return promise.