We build things you can actually watch, click, and break.
AI tools. Game engines. Interactive charts. And sometimes a Madden-style ring under a quarterback, because apparently that needed an entire computer-vision pipeline.
This is https://t.co/sI9yIQQaIv and the Satellite team.
Follow for working demos, the awkward problems behind them, and what we learned fixing them. Plus the occasional developer disaster.
Try the engine atlas: https://t.co/v6gD3uRaoF
See the chart demo: https://t.co/bnEAly9Q4X
Start with the football clip below. #SatelliteAIio
The first real project for our newly rebuilt Rust-based ODIN AI engineering system was something just for fun: building a professional, AI-powered football highlight production pipeline for my son.
What started as "let's make his quarterback highlights look like something out of Madden for the hell of it" turned into a great test of v2 (well, v1 superceding v4 whatever that counts as) Odin.
And I want to give credit where it's due, both to ODIN and to the incredible open-source projects that made this possible.
Here's the technology stack:
- @ultralytics YOLO11x: Detects players in every frame.
- BoT-SORT: Tracks individual players across frames, including appearance matching.
- @OpenCVLive: Camera-motion compensation, optical flow, and image processing.
- @roboflow RF-DETR Seg 2XL: Generates pixel-level player masks. Seriously badass.
- @PyTorch + @NVIDIA CUDA: GPU-accelerated AI inference.
- @FFmpeg / FFprobe: Video processing, clipping, audio, compositing, and encoding.
- @numpy_team + Pillow: Custom geometry, graphics, shading, and alpha compositing.
- @playwrightweb + @ThePSF Python: Hudl footage retrieval.
- @NVIDIADeveloper NVENC on @awscloud EC2: GPU-accelerated final rendering.
- SatelliteGEngine, @reactjs, TypeScript, and @fastapi: The editing interface, playback, and media services.
Now, you might look at that list and think Odin's factories just glued together a bunch of open source but unfortunately ...not that simple. (Continued)
#AIEngineering #ComputerVision #RustLang #OpenSource #SportsTech #Football #YOLO11 #RFDETR #AgenticAI #VideoProcessing #satelliteai
@NSTG_official Mario kart is a little tougher but super Mario World is a great prospect. Would tell you to try out our engine but XR/AR/EM aren't finished π
@lavenderFluffVt By the way, singing and talking activate different parts of the brain - there is a video somewhere of a guy with a speech impediment that sings opera perfectly so don't be nervous just go for it. π₯
What song are you thinking of? I would start with one you know really well, in a key that feels comfortable... you're already nervous about recording, so don't give yourself a ton of extra stuff to learn at the same time. You can do a shorter version for the 800 celebration, too. Record a bit, listen to it, then work on whichever line is actually causing trouble. If it's recorded, you can take breaks and do sections separately. And if your voice is getting sore or hoarse, leave it for another day.
She's already got out three times... I would put her in a closed room before opening the door for groceries or anything else where you can't keep an eye on her. Especially with that road outside. You and your roommate need the same routine. Then teach her to go to a perch away from the entrance and reward her there. Start with just touching the door handle while she stays, then opening it a little, and work up from there. You're giving her something else to do when the door opens, and that's going to take practice. I'd keep using the closed room while you're teaching it. The AirTag is a backup, and it still needs a nearby device to pick it up.
I would separate this into 1) keeping them from overwriting each other's work and 2) keeping them from doing work that doesn't fit together. Separate worktrees help with the first one. For the second, they need the same task list, somebody claiming each task, and agreement on the things they share. Especially with your Blender/Unreal sync... if the exporter changes the format, whoever is working on the reader needs to know that before they build around the old one. I'd have one person or process own merging and deploying, and bring changes together often. Otherwise you can spend days building on different assumptions and have a ton of work to redo when you finally connect it.
What were you leaving it to do? And when it failed, what actually happened? I'd start with one of those runs and look at the last thing it got right... because βthe agent failedβ could mean it lost access to something, misunderstood the job or invented an answer. Those need different fixes. For content work, you could give it three claims to check and ask for the source passage for each one. Let it say it couldn't find support. Then you've got a small job you can check yourself before leaving it alone with a whole pile of research and writing.
I would literally take one photo and make a bunch of versions of it. Too bright, too dark, too warm, too cool... move things far enough that you can actually see what they're doing. Then back them off to something you like. It's much easier to see what a control does that way than move everything a tiny bit and wonder which change helped. Keep the original, though. It's easy to get used to the edit while you're working on it, so come back later and compare them. When you've got merch samples, you can do the same thing with those, except you've got the actual item sitting there to compare the colour against.
I would stop using the card while you figure this out... don't take more pictures on it, format it or agree to repair it yet. Deleted photos can sometimes be recovered, but putting new stuff on the card can overwrite the data you're trying to get back. Check whether Files on your iPhone can still see them, and whether any already made it into Photos. If they're genuinely missing from the card, I'd borrow a computer and a card reader for recovery. Save anything recovered onto the computer or another drive. You don't want to be writing back onto the same card you're trying to recover photos from.
How much have you got to spend (roughly, in KSh), and what do you want to photograph? That changes the answer a lot... pictures of friends outside, fast-moving sport and a dark concert are pretty different things to ask a camera to do. I would look at used options locally, but compare the full setup, with a lens, battery and charger. Some listings are just the body, so the price you see isn't the price to actually start taking pictures. If you've got a few listings you're considering, send those. We can look at what you're getting for the money instead of guessing what'll be available where you are.
Once you've followed a lesson, try one small thing again without looking. It can be the same exercise with a different column or a different number... just enough that you have to decide what to do yourself. Following along can feel pretty straightforward until the instructions aren't there anymore. Then write down where you got stuck and what fixed it, in your own words. Since you're already posting your progress, I'd use some of those examples. Even something like βI got no results because I was filtering the wrong column.β That gives you something you can actually come back to when you've forgotten it next week.
I would give the query some actual time of its own... it's easy to spend ages editing the book and then rush the thing you're sending to get someone interested in reading it. You've got all this context in your head that the agent doesn't have yet. They need to understand what sort of book it is, who it's for and what the main character is trying to do. You mentioned cutting material is the hard bit. This is where I would be pretty ruthless about it. You can love a subplot and still leave it out of the query. Then check each agent's current requirements, especially since it's been ten years. And put the ending in the synopsis. That's where they need to see how the story actually turns out.
You've got an entire world to talk about with Umae... I would use that. Pick a creature and talk through why you made it the way you did. What did you change? What didn't work? Someone who doesn't draw can still get interested in that. And you can get a few posts out of the same piece (without having to finish three more pieces just to have something to post). Henning's offered to share your work too, so take him up on it. I would try a bit of that and see what people actually respond to.
Thanks for trying that. The green flash suggests the Mac starts the session but the stream fails. Next I'd plug the iPad in with a USB cable, tap Trust, and connect again. If wired works, it's the wireless link: turn off any VPN or firewall and restart both. Which macOS and iPadOS are you on? This is where it gets finicky π§