Then during the weekend I saw a pigeon flying into the green ivy on the wall between neightbour’s and mine windows. I thought it could be a nest there. And next Monday I decided to hack a nest streaming project.

GoPro on a stick

Consulted with a LLM, I decided not to proceed with raspberry pi and different pi-cameras, but use my GoPro, which runs GoPro Labs. Starting with a scouting, I’ve managed to find a pigeon sitting on a nest just by using GoPro on a stick and bending out of the open window. Next I collected verbal agreement from my neighbours, that they are ok with me pointing camera close to their window. And started the development.

I was using oh-my-claudecode aka OMC for some of the projects already, so there was no reason not to use it here too. GoPro can stream live to a custom rtmp:// URL. The first design was to run a MediaMTX media server in k8s, stream to it from GoPro and put a static web page with WebRTC/HLS video player, sourcing from it.

After some test runs with Quick app (GoPro original mobile app), I’ve made a QR code that automates GoPro: starts live stream to pre-configured RTMP server, using pre-configured WiFi settings, keeps the stream for 1hr, then stops it, waits 2 seconds, and repeats the automation.

I decided to show that QR code (it does not contain streaming url and wifi info) on https://birds.aleksandr.vin when the stream is offline, so I can point camera to it and start live stream.

It’s recommended to remove battery from the camera when streaming for long time. I’ve played with streaming bitrate settings but didn’t come to the point of configuring a boot-script automation, so that camera can kick live streaming after power failure (I learned about that a day later).

Another problem was the radio signal between camera and home router: it sits on the ground floor in the center of the house, and camera is on the 1st floor on the wall outside of the window. Signal was very unstable and was lost too often, so I brought a WiFi bridge and put it close to the window.

Finally I started the live streaming with a QR code, then bolted a gopro-on-a-stick to the wall while live streaming. The pigeon flew away: it appeared to be very close to the nest, around 20-30 cm. There were one egg visible in the nest.

Whole rigging was a real-life PoC.

Recordings

The bird came back in an hour. I was very glad that she did. I shared the link with friends. When the camera lost the wifi signal and was not picking it up (till the end of the streaming hour), they asked if it is over, and if I am recording it to review later.

So next day I added automated recordings of 10 min chunks, which were uploaded to backblaze bucket and were accessible from a generated list on html page. And in the morning, while taking a train, I reviewed them and found a second egg become visible in the nest!

More networking

The camera was dropping the stream. It could be because the wifi bridge was struggling to keep the radio between camera and home router. I’ve relocated a wired wifi repeater on the 2nd floor to be just over the bridge (which was on 1st floor). Now the route was:

GoPro )))((( wifi bridge )))((( wifi repeater <===> home router

It became a bit more stable since then.

Analyzer + telegram

When stream was live, it was recorded by a special pod and stored in bucket. I kept an iPad by my monitor, where the stream was played. But I wanted to get notified on interesting events, as most of the time the pigeon sat without movement. Next mod was an analyzer pod with some OpenCV computations at 1 fps. You can select region of interest aka ROI for the bird movements and presence and a separate one for eggs. Then a 5 second span was analyzed for meaningful movement of edges. That produced a ‘movement’ event. Bird’s ‘present’ and ‘gone’ events were produced by presence of edges in the ROI. Events were published into a new telegram group and to OTEL.

It appeared more fun to see when analyzer was emitting events, as most of the time they were just noise.

Later I switched to the color-based detection: bird is detected by grey volume, while green, brown (nest) and blue sky sparks were cut off. It improved detection a lot. Later I added a darkness gate: so detections are paused if the amount of color on the frame is too small. And a chain logic: eggs are not detected if bird is present and movement is not checked if bird is not present.

Grafana

Next was building of Grafana dashboard. I’ve used their MCP server to give claude code (or codex) access to my Grafana, and got a pretty decent dashboard (after my review iterations).

Local monitoring

At some point I decided to rig a poor-man monitoring of the network devices: ARP pings and ICMP pings to the camera, bridge, repeater and stream status. Added a systemd service on home raspberry pi, which pinged all devices and published OTEL metrics as status and RTT to device.

Youtube live

HTML 5 video player was probably too simple to handle the fragile video stream, it was stuck sometime and not restarting the reconnected video stream. I decided to try re-streaming my stream to youtube. Added another pod, which used ffmpeg to copy the stream as-is.

This is one of the recorded live streams of that time (7h long): https://www.youtube.com/live/mHFozIH_e4A.

Recovery attempts

Somewhere around that time I saw a pattern: at some point, during a day, camera disconnects, and at the end of stream hour it will appear back but with 10-100x longer RTTs, like it is struggling with something. At this point it can run for full hour with such long RTTs, but if you turn off the bridge and turn it on in 5 seconds, camera got all fixed and is live and fast again. So I put the bridge into a smart plug and tried turning it off and on again via Philips Hue, based on the telemetry from local monitoring, gating attempts – you can see the attemps on screenshot with grafana on Bridge Recovery panel.

More Youtube fackery

While playing with youtube, I registered an app to be able to add comments and update the description. That allows you to create a chapters of the video, once the live stream is ended and video is saved, users can skip to a chapter, which we automatically created based on the events produced by analyzer.

Another fun thing was adding an overlay with a fancy star for youtube stream, which was automatically added when a bird was appearing in the frame. This overlay and commenting proof-of-concept demo (with emission of a test event) you can see in this video I submitted to Youtube for app registration:

With this description:

The liveChatMessages.insert call is made automatically when the computer-vision system detects events (test even emitted in demo video). The call currently returns 404 because this endpoint is gated behind YouTube API verification — this is the verification catch-22. The implementation is complete and correct; it just requires approval to become functional.

Storm nights and hard landings

Somewhere around end of June / beginning of July, it was a storm during the night and the nest, eggs, birds and gopro was ok.

But in two days, pigeon landed on the camera (I was really waiting for that moment to happen) and twisted it up.

I used that as an excuse to add the bootable automation, so camera will boot on power cycle and autostart the main automation.

Being a birdwatcher with a tech

One day I decided to add phone widgets (to both Android and iPhone) which will show a snapshot of the stream at 1fps frequency: done.

Another day, when I was driving in the car, I decided that I want to “see my pigeon” on the CarPlay screen while driving. That showed me two new worlds:

1. Android Auto – a nasty mess of constant everyday updates: you need to install that app on you phone and it will be updated almost every day, to support all different Android installations that are in a wild life. If you have a GrapheneOS and allowed that app to be on your phone, that “everyday-an-update” moto will drive you nuts. And on some days your app would work and you will see a 4fps live stream of your nest on the full screen in your car. And on some random days it will just stop working, maybe because a new update to Android Auto… while still working in emulator.

2. CarPlay – a thing that I was able to see only in the simulator, as you need an official Apple entitlement if you want to sideload your debug app to a real (your personal) car, and using a navigation template (the one that can use the most of the screen) to show a “map” which is a snapshot of my nest at any low fps, was not having a chance for such ‘entitlement’.

Focus on recorded material

After some time I collected quite a list of 10 min chunks, so I upgraded the listing page: added recordings player, added live player, heatmap, added chapters editing and hashtags: https://birds.aleksandr.vin/recordings – so heatmap can highlight recordings with chapters (red color) and you can skip directly to the time of the fun moment (chapters are listed beneath the heatmap).

If you find a fun moment too – send the link to me – link has a timestamp, if you didn’t notice, so if you refresh/bookmark/share it, it will start playinf from the same moment ;)

Here is an example of the link: https://birds.aleksandr.vin/recordings?t=28#birds-2026-07-20_04-21-36.mp4 – the chicks are walking like on a broadway there.

Continuous fun

Maybe I forgot something while was I writing this long story here (it took me some evenings and attempts to go up to this point). I’ll update the story if I find out that what I forgot is worth being mentioned here, once I remember it. But here is one more thing that I decided to build for fun: a macos screensaver https://birds.aleksandr.vin/screensaver – another fun project, which ended up streaming my pigeon nest on a mac, when I was not around – that one for the lulz!

Learnings

1. Pigeons wildlife, incubation periods eggs amount, pigeon milk, learning to fly, etc.
2. gopro streaming stability
3. Youtube broadcasting
4. CarPlay, Android Auto
5. And tons of OMC…

Future plans, once chicks are out of the nest maybe I’ll do them one day

I am thinkg of packaging a system as an ad-hoc live streaming with your gopro on labs: you install labs on your gopro, point it to one QR code, that teaches her the wifi, then to another QR code and you’re live, here and now.

So I started with a search for their names, and quickly discovered one long read Fear and Loathing on the Internet, which is an interesting post to read. While reading it I came over this sentance:

Then on May 5th, my good friend Rob Palmer was killed in an unexplained accident in the Red Sea.

And I don’t remember why, but I decided to search for that *Rob Palmer – the name I never met before – and what I’ve found blew my mind. The long read I’ve found is Get It Wrong And You’re Dead published Divernet on March 24, 1999 at 5:16 pm. That one worth reading in full! I’ll quote a joke from the kicker:

Cave divers used to tell a joke about some of their number who died and went to Heaven. St Peter met them at the Pearly Gates and asked what they wanted to find in their perfect Heaven. A wonderful cave system that no one had been in before, they replied.

Soon they were swimming through a magnificent cave. Suddenly, a diver on a big black Aqua-Zepp scooter roared up from behind them, hooked on to their line and shot off ahead into the system.

The cave divers went back to St Peter to complain that they were not the first into the cave.

“Who was that guy with the big black Aqua-Zepp?” they asked. “Was it God?” “No,” replied St Peter. “That’s Rob Palmer. He only thinks he’s God.”

The ones I was surprized (to say the least) of – I’ll quote here too, feel free to read that post yourself:

Tim and Rob each went in armed with a twinset of air and a sling-tank of nitrox 50 for decompression.

…

I took a casual look at Rob’s computer. At first I thought it read “12m”. A chill ran through me when I realised I had misread “120m”!

… dive with a PO2 of more than 2.7 bar, when he had always advocated an absolute maximum of 1.6 bar.

…

When he expressed the mind-boggling opinion that a PO2 of 3 bar was “safe”, he did so in private.

…

The scene was a typical deep-water reef in the Giftun Islands. It was Jane who, during the descent, decided to stop at 70m. Tim and Jean-Michel say they stopped at 107m, … Rob was last seen below the others, apparently waving them to continue on down.

… within days of Rob’s death, Mila went diving with me and chose to ignore our pre-dive chosen depth limit. I parted company with him at 65m and he surfaced a good time after me with 120m logged on his computer.

And last Friday, the 13th, we executed that idea. He played bass guitar, I played tenor sax, and we recorded jam session on gopro.

A little bit EQ in DaVinchi Resolve

and here we go on bandcamp:

It was a lot of fun and I already want to do it again!!!

All means of reaching the actual help is covered/buffered/greased/hidden and poisined. You will see links or buttons that “suddenly” do not work.

FAQs that you need to steer through to find out an answer how to contact their support (which will just open another door in that labyrinth). FAQ generated by “AI” to “help you” – another layer of dust and bullshit to distract you. Chat with “support” is just another blanket to bump you off the actuall support (if it even exist for that company). It will be a “smart” “AI” bot that will rip your soul asking stupid forms and answering obvious things. Like my recent experience, when PostNL app messages you “Your post is delivered to you and is in your brievenbus!” – I was at home the whole day and at the moment of the message too, nothing was actually delivered and nothing is in my door’s post-hole. You check that post’s track-n-trace and it says “Delivered 10 min ago and post in your door” – quite stupid. You go to the PostNL “support” and end up in chat with “AI” bot. You ask for help as a post was not delivered to you, he asks for a track-n-trace id, and then replies to you like nothing happened: “Delivered 10 min ago and post in your door”. And then it’s a scenario of a phsyco movie: you talk to a dumb asshole, who holds your fate in his hands and nothing moves to resolution.

Support via ServiceNow tickets – “nice one”: you click on “Create a ticket” then you are on a page saying that first you’ll be tried to help by their “AI” – which is a dumb search over their dumb collection of beuracratic instructions… and it finds nothing about your problem. Fuck yes, because I know what is the problem (need to reset a cache in Nexus repository proxy but I lack permissions and is trying to find someone who can do that) but I need to go through “AI” help first. After “AI” surrended I am able to fill in “more details” and a “link to the problematic page” (which does not pass the validation due to presence of fragments in the link, Merry X-mass) and then… good luck to find that small button “submit” (which was initially on the page and was just non-valid until) you pass that torture with “AI”. Ah, forgot to mention – the snowflakes animation on that “support” page was amazing :sarcasm:.

I believe I can easily find more examples of “support” bullshit we are diving deeper with every day.

https://woordenlijst.org

https://nt2taalmenu.nl/nt2-b1-grammatica-menuverledentijd/

https://thedutchonlineacademy.com

TL;DR; I can write posts here with blocks like:

{ % texcret %}
This text must be hidden from public eyes (read encrypted).
{ % endtexcret %}

(I put extra space in tag opennings to go under the radar with the example above).

And they are texcreted into blocks like this at pre-commit via a texcret cli:

WUtMQjIAWoR+A+pe1UDbfMDaAAMtLS0AAeeq6Csra2XDS9x+sKSkQpZACVKqCLx6ngbV9BYAMC83eSK29gi7IOBJ2ryrO1UXZskCpjyCG7i/SL7wYA7+MkHTmz+NwrRdQjFM/d6XPqNIBJywYy2/fbElDne0RGqguuZsEXnFmwPfn3uw63j/UahX3+dGIuWIwFWj6W1Foz+5aJx3xnoK6KQ0qFfdZpCcC/wJILEuYzMpk1HN

Which I can temporarily decrypt (in-page without any network communication) magically by performing the following click pattern on it: dblclick - click - dblclick.

Check the demo or github repo.

Since beginning of August I got locked on a mechanical wrist watch. As a message of disgrace towards the smartwatches, smartphones and third-party apps running on my computers.

How so? Being a computer geek all my life I am loosing fate in them. But thinking deeper over that, I think better would be to say I’m loosing fate in people programming them. Not excluding myself.

I would not brag over the LLMs and all the “AI” things happenning in recent years. Talking in general about computer programms that are here to help people: are they?

You can say — “If it does the job, it’s good enough”. I disagree: consumerism is already leaving piles of crap at places of the “least concerned”. Places where beneficiars are not living. Places where their sellers are not living. Places where those who are living, will get that crap and nothing more.

So apps are everywhere now in our lives. And they are mostly crap. Name it.

Spotify — now playing, while I’m typing it. Just a few seconds ago it skipped a song midair like nothing happened. It’s discovery playlists are useless, I feel hopeless everytime I try them. “Best (something)” playlists are even worser: I hop-off after I exercise the “skip” button for almost a minute, not even being able to listen for some decent time for a song. Should it be matching my music taste or pumping me with some algorithmic one’s?

Proton Mail — (maybe you didn’t even hear of that one) — recent update changed the UX of the archive action, the most-used one, and it feels like a hiccup now when you slide email left.

Guardian App — with heir cursed limit of 30 topics you can follow: I’ve developed that number to almost 50 when suddenly it stopped working and asked me to decrease my choices until I can add new. It is very frustrating to try to find out what is the limit when you stay at 49 and, you remove more and it still says you are above the number, but what number it does not bother mentioning. Later they fixed that by mentioning the limit is 30. Still lame.

Calendar iOS — knows better how to add new event’s info than you: even when you’ve already added the date and time and then picked up a similar-called event from a drop-down list, it replaces you entries with the time of the event you selected for the f#cking name!

Oceanic dive app for iOS — these motherf#ckers managed to release a new version last year that crashed on iOS 17 while still being listed as min required in App Store. There is nothing more upsetting than a crashing dving app on your wrist, when you’re already in a wetsuite in a water set for the dive. Consumer smartwatch from the company with zero experience in diving and an app from the company with zero experience in a app development.

Quick (GoPro) — search the internet, it’s a PITA to use it, don’t want to talk about it more.

Instagram, Facebook — I deleted them at the very beginning of the year and do not miss at all. Could be the ones professionaly built, but you know what — they somehow decided they know better what their users need. Trivial things, they all “do the job” for them: feeding me with their POV of news and posts, algorithms and moderation rules.

Strava, Garmin, Suunto… all busy with their cash flow and not the customers. Check the recent Strava suite over Garmin. Poor Strava was asked to attribute the data uploaded from Garmin devices and they call it a “user’s data” but they did decide themselves what is better for the user when they restricted their API use a year ago.

I can continue the list. But if we go to the professional development, it’s the same: opinionated choices can be good. Can be bad. I meet the bad choices too often. Maybe I just notice the bad things…

So if you can program, do yourself a favor: write programs for yourself.

PS. I wear a mechanical watch on my wrist now and it reminds me that there are good things in life that run no shit and that are completely yours once you bought them. Good luck.