OptionBotZERO is my fun research project. It's an options trading bot that runs strictly on paper: simulated money, but operating against the real market in real time. Real quotes, real market volatility, and a paper broker account taking orders. The only thing that isn't real is the cash.
I built it to learn one fundamental truth: where AI is actually strong, and where speed and plain deterministic rules beat it.
That question matters far beyond trading. It is the exact same architectural principle behind every automation we engineer for impact window dealerships.
The First Plan: "A Council of AIs"
The original blueprint was grand. A Windows machine would rent the smartest AI models on Earth by the penny (Claude, GPT, Gemini, Perplexity) and have them vote on options trades as a consensus council in the seconds the market moved.
It made a great story. Then I audited the actual runtime code.
What the Audit Found
When I measured real execution speed, the findings were humbling:
- The bot waited 30 seconds by default between checks, before adding network latency and token generation time.
- Position lookups, news sentiment queries, and option-chain lookups ran sequentially, not in parallel.
- Every journal write sat directly on the execution path, slowing trades when milliseconds mattered.
None of that was the AI’s fault. But it showed me that asking an AI for its qualitative opinion in the middle of a fast-moving transaction adds delay at the exact second delay is fatal.
The Rebuild: AI Prepares, Rules Execute
The rebuilt architecture split the responsibilities:
- AI works in the background: It synthesizes context, analyzes news, and prepares testable scenarios before the market opens.
- A small, deterministic engine handles the hot path: It monitors price triggers, verifies volume, validates spreads, and executes immediately. Same input, same answer, every time.
Four Rules for Any Automation System
- Let AI think ahead, not in the hot path. Use it to prepare, summarize, and categorize. Let plain deterministic rules handle the live execution.
- Measure before you believe. Every speed claim in an architectural diagram must be checked with real timestamps.
- Count every cost. An automation that works but costs more in token compute and subscriptions than it saves is an operational failure.
- Make failure immediately visible. The goal is never a magical black box; it is an open system where every failure can be spotted, repeated, and fixed.
That is how I build CRM automations for fenestration dealers: plain, deterministic rules that fire instantly, with smart AI handling the preparation.