How I Built a Production System Without Writing Most of the Code

Okay so here is the thing that most people in tech do not want to hear. I did not write most of my own code. There I said it. The V7 trading engine, all 43 modules, the validation framework, even this website. Built primarily through AI agents.

But not these articles. These articles I have saved up and shared with my colleagues, acquaintances, and company quant communities because I like writing. I enjoy it. My mind just cannot stop spilling words out of creativity and ideas. I literally cannot sleep without noting down things that bother me. Of course there is help from AI at a bare minimum, making sure I do not have typos and awful grammar mistakes across the many articles I have written over the years. But I still do not trust AI to fully revise what I want to say without doing proper prompt engineering, so I would rather just get whatever I have thought straight into this, think about it for an hour to make sure it is all good, then check back the next day with fresh eyes. That is the process.

But before you close this tab thinking I am a fraud, let me explain why that actually makes me better at what I do, not worse.

Pre-Beta ChatGPT Days

I got whitelist access to ChatGPT back in late 2022 before it went public. At that point AI coding was rough. Really rough. You would ask for a function and get something that was syntactically correct but logically wrong in subtle ways. But even back then I could tell this was going to change everything.

I started using it for simple stuff. Data processing scripts, basic visualizations, boilerplate code. The key insight I had early was that AI is really good at implementing things you already understand. If I knew exactly what a function should do, the AI could write it faster than me. But if I was vague about the requirements, I would get garbage.

The Architecture Problem

Here is where "vibe coding" gets misunderstood. People think it means you just tell the AI "make me a trading system" and it spits one out. Absolutely not. What it actually means is:

1. I design the architecture. 3 layers, what each layer does, how data flows between them.
2. I specify each component. "This module takes OHLCV data and outputs a Hurst exponent estimate using the R/S method with a minimum window of 100 bars."
3. AI implements it. Writes the actual Python code.
4. I validate it. Does the output match what I know the Hurst exponent should be for this data?

The hard part, the part that takes years to learn, is step 1 and 2. Knowing what to build. Understanding WHY a Hurst exponent matters for regime detection. Knowing that you need cluster-specific L1 thresholds because FOREX and CRYPTO behave differently.

What AI Cannot Do

AI cannot tell you that your feature engineering is wrong because you are using lagged returns without accounting for overnight gaps. It cannot tell you that your backtest is biased because you are using adjusted close prices with survivorship bias. It cannot tell you that a certain threshold works for FOREX but you need a lower one for CRYPTO because crypto has more noise.

That domain knowledge, the quantitative finance understanding, is what makes the system work. The code is just the implementation. Important, yes. But not the differentiator.

The Multi-Agent Workflow

Now I work with multiple AI agents simultaneously. One for code generation, one for review, one for documentation. It is like having a team of junior developers who are incredibly fast but need clear direction. The productivity is insane compared to doing everything manually.

My typical workflow:

This is how I built 43 modules in four years while also learning quantitative finance from scratch. Without AI, I would probably have like 5 modules. Maybe.

The Honest Assessment

Am I a "real programmer"? Depends how you define it. I am not fluent in Python. But I can read most of it by logic, flow, and structure, by visualizing what the code is doing, built up over time from years of coding alongside AI. I can debug code. I can identify logical errors. I understand algorithms and data structures conceptually. But could I write an LSTM from scratch without AI? Probably not efficiently.

What I am is a system designer and domain expert who uses AI as a force multiplier. And I think thats actually the future of software development. People who know WHAT to build using tools that handle HOW to build it.

The 0.08% breach probability does not care whether a human or an AI wrote the code. The Monte Carlo simulation does not check git blame. Results are results.

Honestly Though

Here is the thing I keep coming back to. The instinct when someone questions your coding ability is to get defensive. "I AM a real developer." But the better question is: what actually matters here? The answer is the system works. The validation is rigorous. The risk management is sound. Whether I typed every semicolon myself is irrelevant to the outcome. I would rather be honest about my process and have a system that passes Monte Carlo than pretend to be something I am not and have a system that does not. Update the definition. The world already has.