Sky Parrot, City Lungs

I taught a pocket computer to stare at the sky and guess the air. That’s the joke. A camera points up, a tiny neural net squints at color cast, haze, and cloud edges, then blurts a number like 162. Not because it smelled soot, because it remembers twelve thousand skies from India and Nepal and thinks yours looks a lot like number 8,394 on a Tuesday. Useful, not mystical. More parrot than prophet. And yes, it runs on an ESP32 S3 snugged inside a Unihiker K10, which is basically a lunchbox with opinions and Wi-Fi. It works, somehow, like most things we love that shouldn’t. Think of it as a breathalyzer for the horizon, except it doesn’t blow, it looks, which sounds backwards until you remember the sky wears its bad habits in its colors.

Wait, you already have the question. Why build a sky-guessing gremlin when a PM2.5 sensor costs fifty bucks. Because cities are huge, budgets aren’t, and cameras already sit on rooftops and poles like pigeons with union cards. If you can squeeze signal from pixels, you can cover ten times the map for the price of one sensor and a little gumption. The obvious is still true, the sky goes odd when smoke rolls in or smog sits heavy, and a model can spot big swings quickly. It’s not a lab instrument, it’s a scout with binoculars calling the ridge before you hike into it. Also, training a camera to notice haze changes gives you a second channel of truth when a single sensor hiccups, because redundancy is how you stop a bad day from becoming policy (ask any ops team that lost a weekend to one flaky probe).

Now the part where i level with you. This thing guesses. Politely, with math, but still guesses. It learned skies from two countries, with their light, humidity, dust, cameras. Then you ask it to work in Phoenix or Paris after a sandstorm, at sunset, with a greasy fingerprint on the lens and a camera that white balances like it’s auditioning for a toothpaste ad. It will get confident about the wrong thing. AQI is a cocktail of pollutants, and not all of them tint the sky in helpful ways. Some do not tint it at all. A tiny model that fits on the ESP means shortcuts you will not notice until you do. That is not sin, that is physics. Temper your expectations, hydrate your calibration, and keep a settings log for the particulate-ularly curious.

Here’s the gut-punch. We do not replace calibrated instruments with a camera and a gut-check, then tell certain neighborhoods to breathe easy because the dashboard looks pretty. If a city swaps maintenance for marketing, that’s not innovation, that’s neglect in a new font. Under call a flat gray noon, over call a red sunset, brag the network grew four hundred percent, and the bill lands in lungs. Not only in communities already parked beside refineries and freeways. In yours when wildfire season shows up, when demolition throws dust, when overnight exhaust sits under a low ceiling. Good systems protect everyone because weather does not ask for your ZIP before it settles in.

I like this build anyway, and i’ll tell you why. Used right, it is a flashlight. Early warning, triage, a nudge that says put the runners inside and shut the gym doors. Pair it avec a cheap PM2.5 and a CO2 NDIR, calibrate to a few solid reference stations, retrain monthly with your own skies, and you have a scrappy outpost that hits above its weight. Publish uncertainty, not just a smug single number, annotate failure modes in plain language (you know, like this one) before someone breathes the fine print. Keep a handful of gold-standard sensors alive and public, then show the match between them and the camera guesses in real time. If that sounds annoying, same. Do it anyway. Honesty saves more time than it costs.

Also, the lesson scales. Models repeat their diet. Feed them one kitchen, they cook that recipe everywhere until you teach them new food. Treat them like oracles, they start writing rules. And rules do not cough when they are wrong, people do. So let the ESP do its brave little routine, clap for it, then remember what it cannot see. Teach it your sky. Give it friends. Keep a grown-up in the room.

Actual solutions look different from performance art. Here’s how you separate people who clean air from people who polish dashboards. Cleaners pay for calibration, upkeep, and public data you can check without a login. Polishers print banners, add map pins, and hide caveats three clicks deep. Cleaners publish error bars and fix drift before press day. Polishers point at a big number and rename the problem “progress.” If the budget moves from maintenance to expansion while accuracy slides, that’s your tell.

If you’re still here, you probably want the how, the guardrails, and the civic fine print. Good. Grab your coffee, i’ll grab the solder. Short version next, avec fewer adjectives and more parts you can actually build (ADHD brain insists we’ll still let a few slip through, but smaller ones, promise). Teach a lunchbox to look up, but tie it to instruments that actually taste the air. Celebrate the clever, then fund the boring. The boring keeps us breathing.

--Doc (yeah I said it!)
find me anywhere under @DocAtCDI