Part 2.
This is where the “AI built the product” version of the story becomes slightly incomplete.
Yes, most of the implementation was done with Codex.
No, that was not the hard part.
I also added a full AI telemetry ledger across providers so I could compare:
OpenAI
DeepSeek fallback
latency
provider drift
error states
Debugging based on “it feels different” is not a long-term strategy.
This is a 3-part build thread.
What started as “I’ll just build this with Codex” turned into a full AI engineering system with staging gates, MCP orchestration, telemetry, and an amount of time spent in Railway logs that I’m choosing not to quantify.
Only after that did I start building most of it with Codex.
And yes, the code moved fast.
The system around the coding loop did not.
Part 2 is where it gets more interesting:
MCPs, agent boundaries, staging vs production, and why AI engineering is mostly orchestration.
I’ve been building something called Career OS.
Career transitions are still mostly managed with scattered tools and intuition.
Career OS is an attempt to turn that process into a system.
https://t.co/B6dSeHWJCE
@elonmusk If this is actually true and you plan to stay at the same pace of around 10K new chargers a year, hire me to run this team. I can guarantee it’ll be the best cost effective decision you make. 🙋🏾♂️
23 Model S LR owner. Love the software and driving (sometimes in insane mode). Wish there was an option to browse through or play a slideshow of pics from my phone on the rear screen. @elonmusk@Tesla
I used to be frustrated that while the future seemed so clear to me, no one around me listened or agreed 😤
anyway now that I’m an investor idgaf what other people think, I just get to sit over here making money being right about the future, it’s so great 🤗
Autonomy and responsibility are two sides of a coin and are an important part of the culture of any early stage startup. #culture#startup https://t.co/QRlSHqSQaz