Carlo Lepelaars
@carlolepelaars.bsky.social
100 followers 140 following 87 posts
https://carlo.ai | {AI, Investing, Reading, Music}
Posts Media Videos Starter Packs
Pinned
carlolepelaars.bsky.social
Introducing irouter, a super simple interface to access 100s of @openrouter.bsky.social LLMs with 2 lines of code.

It supports multi-modal inputs (text, images, PDF, audio) and tool loops. Mix and match as you see fit.

pip install irouter
github.com/CarloLepelaa...

Examples (1/6) 👇
carlolepelaars.bsky.social
Released a notebook on how to use Google's new Gemini 2.5 Flash Image for SOTA image editing with irouter.

Simply input image(s) + text and get an image back.

Hope you have as much fun working with this as I have!

github.com/CarloLepelaa...
irouter/nbs/applications/img_edit.ipynb at main · CarloLepelaars/irouter
Access 100s of LLMs with minimal lines of code. Contribute to CarloLepelaars/irouter development by creating an account on GitHub.
github.com
carlolepelaars.bsky.social
For those curious here is the throughput and costs for each model.

- GPT-4o with web search was the most expensive model.
- LFM 40B MoE was the fastest.
- GPT-5-Chat managed to give a great answer without web search.

pip install irouter
github.com/CarloLepelaa...

(3/3)
GitHub - CarloLepelaars/irouter: Access 100s of LLMs with minimal lines of code
Access 100s of LLMs with minimal lines of code. Contribute to CarloLepelaars/irouter development by creating an account on GitHub.
github.com
carlolepelaars.bsky.social
Max Verstappen technically was ahead because of 1 more race win, even though points were tied.

In this example GPT-4o with web search, GPT-5 Chat got the nuance correct. Opus 4.1 realized there was a tie.

Other LLMs were wrong or failed to add the nuance.

(2/3)
carlolepelaars.bsky.social
Added parallelization to irouter. This allows you to do rapid evals for many LLMs at the same time. Your only bottleneck is the slowest LLM.

For example, check 21 LLMs tackle this tricky question from
Formula 1 2021 where Max Verstappen and Lewis Hamilton were tied on points. 👇
(1/3)
carlolepelaars.bsky.social
Just uploaded an example notebook on getting structured output in irouter.

Passing a schema to `response_format` is sufficient to enable structured output.

OpenAI's GPT-5 Mini seems to work well and supports multiple modalities.

More info:
github.com/CarloLepelaa...
Reposted by Carlo Lepelaars
lightonai.bsky.social
After reaching maturity, the PyLate library has its paper accepted at #CIKM2025
🎉Congratulations to the two creators of the library @nohtow.bsky.social @raphaelsty.bsky.social

📄 Read the paper: lnkd.in/eh6NQj4B
Discover PyLate : lightonai.github.io/pylate/
🌐 Learn more about LightOn: lighton.ai
PyLate
PyLate 🐶
lightonai.github.io
carlolepelaars.bsky.social
Adding web search to your LLM app is super simple with irouter.

Chat will automatically keep track of web citations.

pip install -U irouter

Example notebook for web search in irouter:
github.com/CarloLepelaa...

Docs page:
carlolepelaars.github.io/irouter/web
carlolepelaars.bsky.social
Documentation and example notebooks can be found in the Github repository. Feedback and ideas are always welcome. If you are interested in contributing check out the contribution docs.

pip install irouter
github.com/CarloLepelaa...

(6/6)
GitHub - CarloLepelaars/irouter: Access 100s of LLMs with 2 lines of code
Access 100s of LLMs with 2 lines of code. Contribute to CarloLepelaars/irouter development by creating an account on GitHub.
github.com
carlolepelaars.bsky.social
irouter is built on and has the most solid support for
@openrouter.bsky.social, but also offers experimental support for other APIs compatible with the OpenAI spec, like XAI and Moonshot AI.

If you insist on using a different API provider change the base URL.

(5/6)
carlolepelaars.bsky.social
Tool usage in irouter is as simple as passing a list of functions to the `tools` parameter.

This example uses a tool to get the time for a specific timezone.

(4/6)
carlolepelaars.bsky.social
Multimodality is becoming commonplace in LLMs, so we need a simple interface to work with them.

irouter supports:
- text
- image URL
- local image path
- pdf URL
- local pdf path
- local audio path

Just provide a list of strings with the modalities you need.

(3/6)
carlolepelaars.bsky.social
irouter also offers a way to process multiple LLMs at once. Just provide a list of model slugs and irouter will process multiple models. Output is given as a mapping from model to message.

(2/6)
carlolepelaars.bsky.social
Introducing irouter, a super simple interface to access 100s of @openrouter.bsky.social LLMs with 2 lines of code.

It supports multi-modal inputs (text, images, PDF, audio) and tool loops. Mix and match as you see fit.

pip install irouter
github.com/CarloLepelaa...

Examples (1/6) 👇
carlolepelaars.bsky.social
Musy is a work in progress. Short term plans include:
- Tritone sub support
- Scale comparison
- Better chord recognition
- Sheet music generation
- A chord Progression object.

It's my intention to build music web apps with musy.

Feedback and ideas are welcome! (6/7)
carlolepelaars.bsky.social
Chord and Scale objects can be exported to a table for further analysis. (5/7)
carlolepelaars.bsky.social
Chord objects can be extended and transformed with easy method chaining. This includes adding upper extensions, removing notes, inversion and transposition.

carlolepelaars.github.io/musy/chord.h... (4/7)
carlolepelaars.bsky.social
Scale objects allow you to get all modes, chords and secondary dominants from any scale.

carlolepelaars.github.io/musy/scale.h... (3/7)
carlolepelaars.bsky.social
The basic building blocks are Note, Chord and Scale. Each object can be listened to and visualized on a piano or fretboard.

carlolepelaars.github.io/musy/viz.html (2/7)
carlolepelaars.bsky.social
In the past few weeks I have been working on a Python library for analyzing, listening to and visualizing music theory. The goal is that you and I can easily build music-related apps from fundamental blocks.

github.com/CarloLepelaa...

Here are some of the highlights! 👇 (1/7)
carlolepelaars.bsky.social
This overview of AI from a business perspective is the best I have seen!

Definitely also worth a read for AI engineers. There are so many pitfalls to avoid in order to effectively ship AI products.

By @hamel.bsky.social and @gregce.bsky.social:
ai-execs.com
AI Essentials For Tech Executives
ai-execs.com
carlolepelaars.bsky.social
Reading up on recent papers in language modelling as well as some of the classics.

I think the main conclusion is that Noam Shazeer is a genius! 😁

Check out this phrase from the GeGLU paper. 😎

Link:
arxiv.org/abs/2002.052...
carlolepelaars.bsky.social
Wow, this YouTube channel called 2swap has incredible videos which combine mathematical depth, beauty and deep insights!

Check out this video on extending Mandelbrot and Julia sets:
youtu.be/Ed1gsyxxwM0
Mandelbrot's Evil Twin
YouTube video by 2swap
youtu.be