by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
Use Lovable apps at work, safely
A few years ago I built an internal tool in Power Apps for a bank in the UAE. The platform fought with me every step of the way, but it was enough for an internal tool. And at a bank, governance is way more valuable than being shiny and invisible to IT.
This was the trade-off that "citizen developer" always had: limited tools to keep it safe.
AI is also removing these limits in the corporate world. Now someone can describe a tool in the morning and have it ready by lunchtime.
Last Friday, Microsoft moved toward a safer home by launching Copilot Managed Runtime in public preview. Applications created with AI can now run in your company's Microsoft 365 tenant with corporate login, data policies, version control, and many other features a corporation needs.
If your team already builds in Lovable, now it can ship inside your company.
This is the right approach. In a PagerDuty survey of 1,250 office workers at large companies, 89% first used AI tools at home. Two-thirds have already used AI at work they believed wasn't allowed. People aren't trying to break rules, they just want the tools they already know. Now IT can say yes and still have control over what runs.
With AI, we're going to see citizen development pick up.
Why this matters for you: you can finally build the tool you need to save a few hours of work, and it lives somewhere IT already trusts.
I saved Claude usage with Jev
I asked Opus to rename a variable, and it did it perfectly. Probably the most expensive way I could have done it.
I set my default Claude Code model to Opus a long time ago and almost never remember to change it. I bet the same happens to you.
Many of the tasks we give Claude Code are very simple changes like "change a variable" or "add a DB field". The default is set so we keep using the largest model even when we don't need to, and of course this burns more of your usage limit.
I know I should remember to set the model before each task, but it's a lot of cognitive load. The real question is: why don't we have a router that sends the request to the correct model?
This past week, I got a Jev API key, TypeSafe's classification model, and after testing it, it felt like the perfect tool to build a router for Claude Code.
Before writing code, I asked the internet; this time, the internet had already solved my problem. I found jev-router from Pratyush Garg. So it quickly became a buy vs. build decision. Building one meant creating a proxy that depends on Claude Code's internal request format (that can break with any update). Using his meant trusting my traffic to a project that was a week old.
I read the code, and clearly, for personal projects, it was good enough to validate my point. It works as a proxy between Claude Code and Anthropic's API.
For each prompt, it asks Jev whether the task should be answered with Haiku, Sonnet, or Opus. Jev returns a tier with a confidence score, and the router swaps the Claude model before the request is sent. I checked the code, and only the prompt text goes to TypeSafe.
It also has some nice features: when Jev isn't sure, it never drops to a smaller model; for long chats, it never changes the model, so you don't lose the cache, and you can overwrite it for the chat by asking "use opus".
After 5 days, 25% of my requests went to Haiku, so a quarter of my Opus work was really Haiku work.
It's not perfect, since Jev never reads the code; you can pick a less powerful model to fix a bug.
Check your model settings, that one choice is pricing every prompt you send. If you're on Opus by default, try it for a week and check your own split.
On My Radar
The first mass-market agent shipped as its own app rather than a chat feature, and 2.8 million downloads in two weeks says people want AI that acts.
Access to the same model no longer means you get the same results, and the gap between what a model can do and what you can make it do is getting wider.
Token prices halved, but the bigger shift is that OpenAI now sells on cost per task, which is the only number that matters once agents run your workflows.
Agents make thousands of small decisions, and most of them don't need a frontier model, so paying frontier prices for all of them is a mistake.
If your agent reads your email, anyone who can email you can give it instructions, so read this before you connect it to anything that matters
Need another pair of eyes on your AI project?
Thanks for reading.
See you next week. Go build something.
Luis
Know someone who'd enjoy this? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
