CNET’s John Carlsen, writing up his hands-on time with Petlibro’s latest smart feeder, described the standout piece plainly: the version he tested adds an AI-powered camera that can recognize which pet is eating and automatically associate the meal with the correct profile. He called that especially useful in a house with more than one cat.
Read that as a pet owner and it sounds like a convenience feature. Read it as someone who builds bots for a living and it sounds like identity resolution running on a countertop appliance. That is a harder problem than most feeder marketing lets on, and it is the part I actually want to talk about.
Why the recognition step is the whole system
Every automated feeder before this class of device had the same blind spot. It knew how much food it dispensed. It had no idea who ate it. In a single-cat home that gap does not matter, because dispensed equals consumed. Add a second cat and your data quietly turns into fiction. One cat is over target, the other is under, and the log says both are fine.
Petlibro’s answer is to put a camera in the loop and attach each feeding event to a pet profile. Once that association exists, the rest of the feature list stops being a pile of separate toggles and starts behaving like a coherent system. The feeder tracks individual cat meals. It supports multiple diets. It runs scheduled feeding and smart refill modes. It sends alerts when something looks off. None of those work well without a reliable answer to “which cat is this.”
The pipeline, as a builder would draw it
Strip the product down and you get a shape that will look familiar to anyone who has wired up an event-driven service:
- Trigger — a pet approaches, or a scheduled dispense fires.
- Identify — camera-based recognition maps the visitor to a known profile. Petlibro’s Granary 2 line also carries RFID in the family, which is a second signal for the same question.
- Act — dispense according to that profile’s diet and portion rules, or hold.
- Record — write the meal to the right pet’s history, with video you can check later as proof of feeding.
- Notify — push an alert when the system detects an issue.
The interesting engineering question sits at step two, because classification is probabilistic and feeding is not. You either dispense or you do not. Somewhere in that device a confidence score becomes a binary decision, and whoever tuned that threshold made a real call about which failure hurts less: feeding the wrong cat, or refusing to feed the right one.
Where I would push on it
Two things I would want to know before trusting it as my only source of truth.
First, degradation behavior. Cameras deal with bad light, odd angles, and cats that approach sideways. Two similar-looking shorthairs are a genuinely tough discrimination task. If recognition fails, does the feeder fall back to a secondary signal, hold and alert, or guess? A system that tells you it is unsure is more useful than one that quietly logs a wrong profile, because bad data that looks clean is worse than a visible gap.
Second, single-hopper constraints. Petlibro supports multiple diets through scheduling and portion control, but the current lineup dispenses from one food store. A competing design from PETKIT uses a dual-hopper build with two separate compartments, which lets two different foods come out of one unit. That is a physical answer to a problem Petlibro solves in software. Neither approach is wrong, but they fail differently, and the split is a nice reminder that architecture decisions are rarely just code decisions.
What this means beyond the food bowl
The pattern here is the one I keep running into in bot work. Automation without identity produces tidy logs and useless insight. The moment you can attribute an action to a specific actor, aggregate numbers turn into individual behavior, and individual behavior is where the signal lives. A drop in one cat’s intake is a health flag. A drop in total household intake is noise until you know whose it is.
Petlibro’s feeder is a small, domestic version of that idea, shipped as a product you can buy with a 24-month warranty. The camera is not the feature. Attribution is. The camera is just how the device earns the right to say something specific.
If you are building anything that watches a shared resource and reports on it, the design lesson transfers cleanly. Decide early how you will identify actors, decide what happens when identification fails, and be honest in your logs about the difference between a confident match and a coin flip. A feeder that admits uncertainty is doing better work than one that pretends to know.
🕒 Published: