Noon-fiction stories.
Tracking a ball in sports video is kinda a nightmare.
Just pause any padel or tennis stream mid-rally. There's no ball there. There's a blurry smudge. Or two smudges. Or nothing cause some guy is standing right in front of it.
A normal detector looks at one frame at a time and has to go, "ball is here." It's basically guessing. One bad frame, and you lost the ball, and the bounce, and every stat after that.
What worked for us was feeding the model a few frames in a row rather than one. So it kinda knows where the ball was a sec ago and where it's flying, and one blurry frame stops being a big deal.
There's a 2019 paper on this: TrackNet in tennis. ~98.5% F1 on the match they started with, ~84% on other videos. Yeah, that drop is pretty much our whole job.
The first match always looks amazing; then new court, new camera, evening light, and you're back to work.
For years, we basically had one sales channel. Everything else was referrals and a bit of luck.
And it worked, which is the annoying part. When something keeps working, you stop looking at it.
One channel isn't really a sales system.
It's more like something you depend on without noticing, until it stops.
We also sold everything to everyone through that one door. So the channel never told us who actually needs us most – we just said yes to whoever showed up.
And we never recorded a single sales call. Eleven years of what clients asked for and why they said yes, and it's gone. That one still hurts a bit.
So now we're building the rest: outbound, partners, and this account too.
Pretty typical story with clients: they take a good detector; it looks great on benchmarks, and on their real footage it misses half the stuff.
Most of the time the reason is kind of dumb. The model never sees the original image. You shoot in 4K, the model gets something like 640px, and the ball, or whatever small thing you're after, is now 5 or 6 pixels. Not much to detect there.
What we usually do is two passes. First, the model looks at the whole frame, and on the full picture you can often already tell something is there and roughly where. Then we cut the frame into tiles with some overlap, run detection on each tile at full res, merge them back, and that's where the small stuff actually gets found.
There's a 2022 paper (SAHI) where the slicing step alone gave around +5-7 AP on aerial data.
On a server, it's fine. On a phone or a Raspberry Pi, you pay with time; more tiles mean more passes. And that's the part where it gets interesting for us.
For a few years, we were growing, and I was pretty happy about it. More people, more projects, revenue going up.
Margins were going down the whole time. Slowly, a bit at a time every year, so nobody really paid attention.
Money was there, so why worry?
Looking back, those good years were exactly when we should have built the boring stuff – processes, proper analytics, a sales system that doesn't depend on luck.
We didn't – we were busy hiring. I was too.
When the market turned, we found out pretty fast what we actually had and what we didn't.
Spent 11 years selling everything. Web, mobile, backend, design - you name it.
Here's the problem: AI made the "we can build anything" part cheap. A founder with an intern and an AI subscription can ship an app now. So a generalist agency is competing with that.
What AI didn't make cheap is the hard stuff. Getting a model to see tiny objects, run on an edge device with limited resources, and work on real footage.
For us, that's computer vision. We've been doing it since 2019, and most of our biggest clients came through it.
So we're narrowing down. A bit late. Doing it anyway.
The shocking reality is this.
Every student can write code now. Who's responsible for the quality is a different question. "Speed beats quality" works for a demo.
However...in complex systems every skipped check comes back as an incident.
Quality needs checking everywhere, and most of all there.
Thoughts?
Appeared to be a simple scam message.
Developers, be careful!
It just tells you to download the code and review it to make an estimate.
And at the end of the day, it asks you to check out the specific branch for the NDA signing.
When you actually do, the post-checkout hook downloads the malware script, which steals your passwords.
It's crazy.
Hey businesses, how important is it that your data stays within your network perimeter?
I got a very good idea and try to understand if it worth implementation.
Hot take: back in 2022 the problem was follow-up, not leads.
In 2026, it’s flipped. Leads are scarce again, and a generic offer doesn’t cut it.
Buyers need a sharper, more advanced pitch to even bite. The offer that works even when the founder isn’t in the room closing it personally.
That’s why I stay out of client-facing sales. A founder who’s the best closer isn’t a strategy. It’s a single point of failure.
Brainstormed a quick validation for Upwork sales pipeline, together with the team.
We're going to use @GigRadar_io API for jobs' retrieval.
Then to validate through proprietary algorithm, craft a custom personalized proposal with the following auto-bidding.
Is there anyone who did similar things?
Share you thoughts guys.