## Why Am I Doing This? I have been diving headfirst into cryptocurrency investing for a couple of years now. Admittedly, I am no better at it than I was on day one, but hey, I am trying my best! I mostly invest in spot crypto. I hold and slowly dollar cost average my way into favorable positions, usually buying the dumps and then holding on for way too long during the pumps. I decided that just holding was not enough. Driven by a little bit of greed, I started looking into futures, specifically trading with leverage. The rough premise is this: I use my initial investment, borrow some extra capital, and my position gets multiplied. If you guess right (like betting an asset will go up), your gains are massive. But it goes both ways, and the losses can absolutely wreck your wallet. I tried trading futures manually and actually made some cash for a bit. But a couple of brutal losing streaks really messed with my psyche. So, I thought, why not let a bot handle the execution? A bot is awake all the time, needs zero coffee breaks, and has literally no other job. What could go wrong? I was heavily inspired by OpenClaw. I figured I could build a simpler version that just pings my Telegram every time it makes a trade or spots a juicy market pattern. ## The Goal My target was pretty straightforward. I bankrolled the bot with 70 bucks and expected a 5 dollar daily return. The bot was programmed to run until it hit that goal, lost 2 dollars, or maxed out its daily trades. I quickly realized aiming for 5 bucks on a 70 dollar stack was a bit ambitious, so I dialed it down to a more realistic 5 percent per day. The operational budget for all this was around 10 bucks a month, split evenly between a Railway deployment and OpenRouter API costs. Super reasonable and totally achievable! ## The Execution ### Hyperliquid A buddy of mine who built a similar bot told me Hyperliquid was the way to go. You just set up a wallet and connect the API to a script. He was right about it being easy, but he conveniently forgot to mention that their Python SDK documentation is pretty rough around the edges! ### Telegram Getting a Telegram bot up and running was a breeze compared to other platforms. I just registered my bot with BotFather, and boom, we were live. I added a few custom commands so I could manage trades without even opening the Hyperliquid app. I named the bot Babonie Trade Men, inspired by my girlfriend's very cute dog, Baban. His mission was simple: * Guard the stash * Make me some cash * Hopefully avoid licking any dead rats while he is at it (IYKYK) ![Baban the Dog](https://1s82jmgeo0ib5jco.public.blob.vercel-storage.com/blog/babonie-trade-men/a-journey-in-losing-money-IMG_20260302_185732%20%281%29-3409924398d4de958a9f44bcdbc6e03d.webp#dim=400x400) ### Cron vs Crony? The bot's main loop evaluates the market and executes trades on a set interval. I figured 15 minutes was the sweet spot. I also designed a longer Thesis Cycle that evaluates the broader conditions for the symbols the bot is watching. This includes generating custom strategies and backtesting predefined ones. I could not just use a simple CRON job though. Since the app is a single service, the Telegram webhook needed access to the bot's local memory. In hindsight, I totally could have avoided this headache if I had just used a fast persistence layer. Lesson learned! ### Staying Relevant with AI The heavy lifting (decisions, sizing, thesis generation) was powered by LLMs. I basically firehosed market data into the AI and let it make choices for me. I went with Deepseek v3.1 and v3.2 for the core brain. I chose Deepseek mainly because it won the first season of the nof1 trading challenge! I thought about using Grok, but it was just a bit too pricey for my budget. For the Telegram notifications, I used Step 3.5 Flash since it was free on OpenRouter. Massive win! ### The Architecture Here is the refined look at how all the moving parts talk to each other: ```mermaid graph TD subgraph Thesis [THESIS LAYER: The LLM Brain] TA(Thesis Agent
Analysis) ==> MC{Market Context
Regime} MC ==> SS(Strategy Selector
Configurator) end subgraph Orchestration [ORCHESTRATION LAYER: The Engine] CR(Cycle Runner
Thesis & Exec) ==> SigS[(Signal Store
Management)] SigS ==> Rec(Reconciliation
EOD BOD Sync) end subgraph Execution [EXECUTION LAYER: Smart Actions] EA(Execution Agent
How to Exec) ==> OP[/Order Plan
Actions/] OP ==> Tr(Hyperliquid Trader) end TW(((Telegram Webhook
Cmds & Status))) DB[(PostgreSQL
State & Logs)] Thesis ==>|Passes Strategy| Orchestration Orchestration ==>|Triggers Plan| Execution EA <==>|Alerts & Commands| TW Tr ==>|Records Trades| DB ``` ## Did it Work? Well, to put it nicely... not exactly. ### The Downs This project took a serious toll on me, physically and mentally. I was staying up super late fixing gremlins in the codebase, and even then, I never fully trusted the thing. I found myself obsessively checking the Hyperliquid app every five minutes. I even spammed my own Telegram command just to check the PnL. ![PnL Madness](https://1s82jmgeo0ib5jco.public.blob.vercel-storage.com/blog/babonie-trade-men/a-journey-in-losing-money-a-journey-in-losing-money-image-203820ba563b8e7e08e89012d296794f-010d6e484bd0c5419eb1a4c4ba72fd85.webp) In the end, I lost about 36 bucks. Painful, right? I realized I had configured the bot to be way too aggressive. It was taking garbage trades just for the sake of being in the market, which was definitely not my brightest moment. I also struggled to code in proper Take Profit and Stop Loss logic at first. That is a massive rookie mistake because it left my positions completely exposed to risk. I eventually fixed it, but that little oversight cost me about 20 bucks alone. Even after squashing the bugs, the entries and exits were lagging, making profitable trades incredibly hard to catch. ### The Ups It was not all doom and gloom! The bot actually strung together four consecutive green days at one point, which really saved the PnL from completely tanking. I felt like a proud dad watching my digital son finally bring home the bacon. I was also super stoked about the Telegram integration. Opening and closing trades from a chat app was incredibly slick, and getting the automated market reports was genuinely exciting. The bot also learned some patience. There were times when it would sit on its hands for almost a whole day, only springing to life right before the trading cycle closed. It even picked up on some cool market trends, finding a symbol that traded completely counter to the rest of the market (hello, hedging opportunities!). ## The Wrap Up Building an AI trading bot was a wildly fun and chaotic experience. Sure, the execution could have been better, but I paid for this lesson with my own time, effort, and cash. For now, Babonie Trade Men is getting some well deserved rest. But one day? I might just spin him back up. I will leave you with a heavily paraphrased quote from the friend who helped me through this: > "The natural enemy of artificial intelligence is human stupidity."