I’m Jay Derinbogaz. I’ve founded three companies, and I was putting large language models into production three years before ChatGPT turned that into a job title.
I was here before the API
November 2022 is when most people met a language model. I’d been shipping on them since 2019.
OpenAI staggered the GPT-2 release across that year because they judged the full 1.5-billion-parameter model too dangerous to publish. I was fine-tuning the checkpoints as they landed. No chat endpoint, no system prompt, no vendor to call — you scraped your own dataset, you trained it yourself, and you made it fit on hardware that was never designed for it.
NameKrea came out of that. GPT-2 355M, fine-tuned on 100,000 domains I scraped from the Majestic Million and paired with the meta description of what each site actually was, so it generated names that meant something instead of noise. Thirty minutes of training across two GPUs. It shipped as a live product, it’s MIT licensed, and people are still forking it years later.
Every significant open release of that era went across my GPUs — gpt-2-simple,
finetune-gpt2xl, aitextgen on multi-GPU SageMaker, then GPT-J 6B running on
consumer VRAM. I got the full 1.5B GPT-2 XL training on a single 16 GB card when
the accepted answer was that you couldn’t. That was the whole discipline back
then: the models were free and the compute was the wall, so the engineering was
all in how you got around the wall.
The industry now spends billions on the same question I was answering in a bedroom in 2019. What does it take to get this thing into production on hardware you can actually afford?
TextCortex
I co-founded TextCortex in Berlin in 2021 and I’m the technical founder. We build AI agents for enterprises that have to care where their data lives — European infrastructure, real governance, deployed inside companies like Siemens and Lamborghini.
Backed by b2venture, Speedinvest and Entrepreneur First. Over two million users. Named a category leader by Gartner, G2 and Capterra, #20 Fastest Growing Product of 2024 and #10 Best AI Software of 2025.
I still read the diffs. Most of what I build now is the tooling around the models rather than the models themselves, and both of my current open source projects came straight out of problems we hit running engineering at scale.
Juror runs several frontier models across the same pull request in parallel, because one reviewer means exactly one blind spot and you will never learn its shape from the inside.
GitRank solves the other half. It scores every merged pull request with Claude — classifying the work by component and severity, weighting a P1 authentication fix differently from a copy change — and turns that into leaderboards, velocity metrics, bug hotspots, feature ownership and knowledge-silo detection. Engineering performance review is still mostly vibes, and the data to do it properly has been sitting in everyone’s git history the entire time. Self-hostable, EU infrastructure, free for non-commercial use.
Three companies
TextCortex is the third.
At 22 I founded an AR/VR company building for Renault and Nokia. At 24 I founded Grafentek in Germany to make graphene battery materials — deep tech, hard science, and the German business press ran a book chapter on it. In between I led data engineering at trivago.
I studied mechanical engineering with a focus on computational mechanics — numerical methods on very large matrices, which is a far shorter walk to training neural networks than it sounds. I got to language models by way of graphene and virtual reality, and in every one of those fields the hard part was never the demo.
What you’ll find here
Build logs, post-mortems, and receipts. I publish the actual numbers — what it cost, what it caught, what it missed — because in this industry almost nobody does, and the arithmetic is usually the entire argument.
About this site
Static Astro, deployed to Cloudflare Workers. Posts are markdown files in a git repo. Every post is translated into eight more languages by a GitHub Action, including the URL slug — those translations are machine generated and labelled as such.
The GitHub and LinkedIn links in the sidebar are the fastest way to reach me.