HOW APEX IS BUILT
BUILT WITH AI
One person directs the work and AI writes the code: Gemini for Eras 1 to 6, Claude since September 20.
RUN BY RULES
No AI decides what APEX trades. Every decision follows tested rules and statistics.
CHECKED IN PUBLIC
Every settled result posts live, and each day is sealed so it can't be changed later.
ERA 01 · AUG 12 · V3.7 – V4.5 · BUILT WITH GEMINI
A SCRIPT THAT CLICKED BUTTONS
APEX started as a Python script driving a web browser. Every 15 minutes it opened a retail trading page, read the odds on where Bitcoin would close, clicked to place a bet and scraped the receipt to confirm it had gone through.
It didn't take the page's word for who won. A small tracker of its own checked Bitcoin's price every 2 seconds and graded each session itself. When the site stopped it to check it wasn't a bot, it paused and waited.
By the end of that first day it was reading the odds every 12 seconds and no longer stumbled when one 15-minute session handed over to the next. It was fragile, but the loop was there: watch, decide, bet, grade, repeat.
ERA 02 · AUG 13–16 · V5.0 – V5.9 · BUILT WITH GEMINI
A NAME AND A BRAIN
This is where the script became APEX, and where it stopped running on a timer and started reacting to the market as things happened.
It also tried a lot of ideas at once. It bet on prices snapping back from extremes (an RSI above 75 or below 25, or a sudden burst of volume), with a trend filter nicknamed the "Apex Predator" to keep it out of runaway moves. A second hunter, the Micro Sniper, looked for moves in the middle of a session, and signals from the 1, 5 and 15-minute charts had to agree. Daily profit targets, daily loss limits and failsafes for photo-finish closes sat around all of it.
Three things from that week are still at its core: every bet goes into a database, alerts go out the moment something happens, and results come from the exchange's official settlement instead of scraped estimates. The strategies themselves didn't last. By late September APEX had narrowed to a single signal.
ERA 03 · AUG 16–21 · V6.0 – V6.5 · BUILT WITH GEMINI
CHOOSING ITS OWN PLAY
APEX started making its own calls within a session, and starting up with nobody at the keyboard.
A mode called Auto-Macro read the trend partway through each session and switched between betting that the move would reverse and betting that it would continue. At startup APEX pulled the last four settled results from the exchange, so it never began blind.
It also picked up some hard rules. No orders in the last 30 seconds of a session, a rule nicknamed the "iron wall". Ignore any price that would pay more than 25 times the stake, because that usually means a broken order book. And keep running through an exchange outage instead of falling over. Even stopping it changed: a clean shutdown on Ctrl+C, the difference between switching a bot off and crashing it.
ERA 04 · AUG 21–28 · V7.0 – V7.5 · BUILT WITH GEMINI
STRAIGHT TO THE EXCHANGE
The retail page APEX had been clicking through was only a window onto an exchange. So APEX dropped the browser and went straight to the source: no more scripted clicks, just signed requests to the exchange's own interface, reading the order book and checking every receipt.
Within two days it traded more than Bitcoin, adding other cryptocurrencies and commodities. It also kept a live data stream open, so it saw the order book change as it happened instead of asking for it.
The rest of the week was about not falling over: handling markets that vanished mid-request, retrying settlements, keeping its credentials out of the code. It even got cloud licensing, so a copy of the engine needs a valid key to run.
ERA 05 · AUG 28 – SEP 11 · V8.0 – V8.4 · BUILT WITH GEMINI
ONE ENGINE, MANY MARKETS
Until now every market had its own copy of APEX running in its own terminal window. Era 5 replaced them with one engine: a single program driven by one settings file, with an agent for each market sharing one live data feed and one database.
By the end of August it watched fourteen markets, nine cryptocurrencies and five commodities. Before every order it priced the true cost of filling it from the order book, and it learned to split large orders into smaller slices so it didn't push the price against itself. It learned the calendar too: winding down before the exchange's weekly maintenance, and going dormant when commodity markets close for the weekend.
On September 9 APEX moved to its own server in the cloud, where it still runs. One idea from this era, a recovery mode that carried losses into the next bets to win them back, was switched off in October.
ERA 06 · SEP 11–20 · V8.5 – V8.7 · BUILT WITH GEMINI
TUNING ITSELF
APEX became a closed loop. A separate program, the Autotuner, studied recent market outcomes every six hours and rewrote APEX's settings. The engine noticed the new file, reloaded it without stopping, and posted what had changed to Discord.
Two guards arrived. The first Sentinel kept a calendar of big economic releases, like Fed decisions and the inflation and jobs reports, and stood APEX down around them. The second watched Bitcoin, Ethereum and Solana prices directly and held back APEX's next bets if one of them fell 2.5% within four minutes.
Smaller things mattered too. A broken run of bets was now cleared without logging a fake win, so the books stayed honest. And APEX started writing a heartbeat every 2 seconds, the same signal the live page reads today to show whether it's LIVE or PAUSED. This era's last version, v8.7.6, is where the build changed hands.
ERA 07 · SEP 20–28 · V9.0 · BUILT WITH CLAUDE
HARDENING AND VALIDATION
On September 20 the build changed hands. A different AI, Claude, started with a full review of the running engine, and the next eight days came down to two questions: could APEX survive things going wrong, and was its edge real?
Surviving first. Every write to the database became crash-safe. APEX now picked up where it left off after a restart instead of starting over, kept one market's crash from taking down the rest, and stopped tripping over daylight-saving changes. 77 automated tests checked it all, and the dashboard's load on the server fell from about 41% to about 10%.
Then the harder question. An audit of the Autotuner found it approved settings on pure noise 58% of the time, and once it was corrected, the tuned settings showed no proven edge. That could have been the end. Instead, a study of real market outcomes found a small but real tendency, and on September 23 APEX was narrowed to that one signal. Every version since has built on it.