Geoff Cox

@CoxGeoffrey

Hands-on engineering leader. Python, TypeScript, AWS and AI-assisted development. Creator of MSON.

Seattle

redgeoff.com

Joined November 2020

141 Following

33 Followers

Replies Reposts Media

Geoff Cox
@CoxGeoffrey
Nov 3
I stopped re-explaining myself to AI. Now every AI I use reads one git repo of markdown notes first and updates it by pull request. A checker keeps corrected facts corrected, and a weekly "dream" tidies it up. There's a template to start your own:
I stopped re-explaining myself to AI
From redgeoff.com
Geoff Cox
@CoxGeoffrey
Oct 27
Our kids' schools sent ~200 emails in six weeks, plus an app that never reaches email. I built an AI bot that turns it into one digest a day. Getting it to summarize was easy. Most of the work went into deciding what it can never do.
I built an AI bot to read our kids' school email. Here's what I'd tell you before you build one.
From redgeoff.com
1
Geoff Cox
@CoxGeoffrey
Oct 20
21 months into building a trading system, I had two AI models evaluate whether it should become a company. Both said no. Then I spent weeks arguing with the verdict. 25 arguments, none moved it. How I set it up so I couldn't fool myself:
I spent 21 months building a trading system, then had AI models tell me not to start the company
From redgeoff.com
Geoff Cox
@CoxGeoffrey
Oct 13
I ran the same trading model 11 times, changing only the random seed. The best run scored about twice the worst. 14 of the 21 keep-or-drop decisions in my research log sat inside that noise. What I changed:
Changing only the random seed made my trading model look twice as good
From redgeoff.com
1
Geoff Cox
@CoxGeoffrey
Oct 6
Backtests lie in quiet ways. My first live trading models were trained only on companies still in the S&P 500, a fix that zeroed most of a model's features went live for 13 months, and an AI reported leaks that weren't there. What I learned:
My first trading models knew which companies would survive
From redgeoff.com

Geoff Cox

Geoff Cox’s posts