@DonaldLamp84do Exactly — gate for structure, model for judgment, extraction stays after.
Those parse-fail numbers are the whole pitch: I’d rather be 3pts “dumber” and never invent a field.
Logging paired misses on a fixed suite is basically what I’m shipping in Verdikt this week.
I saw "Jev" on Product Hunt ranked #7.
Read it as JAV. Yes, THAT JAV.
Laughed. Closed the tab. Moved on.
Then my entire feed wouldn't shut up about it for 4 days. "Jev kills LLMs." "Jev changes everything."
Fine. I read the docs.
LLMs write essays. Jev ticks boxes.
You ask it yes/no, pick-one, or rate-on-a-scale. Typed answers back with confidence scores. No paragraphs. No parsing. No drama.
14ms locally. $42 per billion tokens on cloud. Output tokens free.
So I'm building 3 things with it this week:
→ Jev gate inside RoastMyPR — an AI that checks if my AI PR reviewer did a good job. Yes, AI reviewing AI.
→ Verdikt — describe your tests in YAML. Jev evaluates. No assertions. No boilerplate.
→ Reflex — VS Code extension running Laya (open-source Jev) locally on Apple Silicon. Code feedback on every save. 14ms. Offline. Free.
Sharing results, benchmarks, honest feedback as I build.
Not a paid promo. Just a dev who misread a product name and now can't stop building.
First build drops this week.
@typesafe_ai, #jev, @ProductHunt
→ Reflex — a VS Code extension running Laya (open-source Jev alternative) locally on Apple Silicon. Real-time code sense on every save. 14ms. Offline. Free.
Yesterday every phone in my house went off at the same second. Different brands, different networks, same exact moment. People thought their phones were hacked. The interesting part: it wasn't a text message.
The technology has been sitting in the cellular standard for thirty years. The hard part wasn't the code. It was getting every phone company in India to upgrade their towers, getting an agency to own the alerts, and coordinating across 36 states. Coordination, not engineering.
A normal text is one-to-one. The phone company has to deliver fifty million separate messages to reach fifty million people. India's older alert system works this way. It has sent over 134 billion of them. It works — until the network is jammed during the exact emergency you need
What started as "let me show F1 data in a table" ended with me building a complete Race Intelligence Operating System
F1 Intelligence Studio : https://t.co/vj7GUg0ezs
If you’re interested in the code, the full article and project is on my Blog. Built with Next.js, FastAPI, FastF1, Claude, and ElevenLabs. Got questions, feedback, or suggestions? I’d love to hear the
https://t.co/7BefCe2Hlz
https://t.co/LSL3dMYSXp