$ whoami
John Mark Bulabos
Builds things. Breaks things. Automates the rest.
Hey! I'm a full stack engineer who builds things mostly to find out how they work. I run the experiments and share the results, including the ones that crashed and burned, so you don't have to repeat them.
I love automation (if I've done something more than twice, it's already a script) and lately most of my experiments involve AI, mainly to find out what it's actually good at and what it just says confidently. Say hi on GitHub or LinkedIn.
// latest posts
all posts →- Paid AI judge Jev lost to free, fine-tuned Laya, 27% to 83%
Free fine-tuned Laya vs paid Jev on 2,000 rated comments: 83% vs 27% matched the raters' majority on 7 toxicity questions, 81% vs 71% with tuned cut-offs.
- Chinese OCR model beat Google's Gemini on price by 8x
Of 14 cheap vision models on 200 OCR images, Ant Group's Ling 3.0 Flash VL scored within 3 points of Google's Gemini Flash Lite at an eighth of the price.
- They said a code graph cuts Claude Code costs. Not on Opus
CodeGraph cut Haiku 4.5's estimated Claude Code cost 27% on long tasks (16 comparisons). Opus 5.5 saw no clear saving, and 30% more when told to use it first.
- Python to Rust: 73% fewer servers for our AI voice engine (in load tests)
In one load test each on the same server, Rust carried 3.75x the calls of our best Python setup after a rewrite, so the same calls would need 73% fewer servers.