[Roni Bandini]’s JEV WiFi Analyzer is an interesting experiment that glues together a single-board Linux computer, a USB WiFi adapter, and a specialized AI model to create a new kind of automated tool that passively monitors WiFi activity to measure and profile coverage, congestion, and suggest possible optimizations.
Here’s how it works: a Python program connects to WiFi, cycles through available channels, and uses scapy to gather details like signal strength, number of clients, WiFi frames per second, and similar data about each given channel. That data is fed to Jev, a highly efficient sort of decision-making AI model responsible for interpreting that data on an ongoing basis.
Briefly, Jev is a new kind of AI that is completely unable to generate text or images. It outputs only numbers representing bounded decisions. We covered these decision-making models earlier, including locally-installed versions. Jev is cheap enough to throw at a variety of problems, but it’s also practical to run your own model on your own hardware.
In [Roni]’s analyzer, Python gathers the raw numbers and Jev takes care of classifying the WiFi network’s coverage, congestion level, and basic remedial actions if applicable. There’s a brief video embedded below (Spanish only) that shows the Jev WiFi Analyzer in action.
The AI isn’t making particularly deep analyses, but it doesn’t take much to imagine Jev taking a more active role like deciding which channels to spend more (or less) time monitoring. Using a decision-making model in this way is interesting, and makes us wonder if similar models integrated with other kinds of hardware is a branch of projects waiting to happen.

Why is it called a LLM in the tags?
These decision-making models are a type of LLM that can understand language as input/instructions, they just can’t generate it.
Why does this need AI?
Youtube clicks
Nowadays, any software that incorporates “switch” or “if-else” structures, or “while” loops, is labeled as AI.
Back in the day, we only had “if”. That was true AI.
So, if I understand this, this is a rig that scans the local Wi-Fi environment, profiles it, and tells you which channel is least congested?
That’s my read. AI seems a very poor solution relative to basic statistical analysis.
Big question is what one would do if the congestion is dynamic. For example, neighbors come home and turn on several WiFi linked TVs, etc.
Is there a way to change your own channel dynamically without dropping connected devices? If there is it would seem a natural feature of a good access point.
That´s what many ( most ? ) reasonably good access points do. Set the channel to Auto, and the thing tries to find a less-poluted channel. Some older equipment ( notebooks, tvs, macintoshes, whatever ) do not like that very much and may have trouble connecting.
As much as I’ve gathered, Jev is a classifier – it’s trained to separate the inputs into categories like “yes/no” by arranging them in some multi-dimensional space and finding how they cluster together.
Classical classifiers are trained on what they’re meant to decide on and they accept no other inputs. Jev has a full language model, so it can understand written questions, and as a result can classify things differently depending on how you ask the question – or by what other information is present at the same time. The answer to 2 + 2 may not be 4 every time, and the model may confidently declare a different answer if it’s told that a bluebird is singing in the woods. It’s an artifact of trying to cluster data with irrelevant information mixed in.
In other words the model can behave like a frightened horse: a stick on the ground that looks enough like a snake may spook the horse, throw the rider and break the cart. That’s why horses wore blinders – so they wouldn’t react to all sorts of irrelevant noise – and why statistical classifier AIs should not be too general for the problem, or they should operate with heavily filtered inputs, which defeats the point of Jev.
I use my A-eyes to read a spectrum scan to find the least congested channels :P
An interesting use of Jev model for a purpose it wasn’t designed for.
Also amusing to read the old COBOL coders moaning about any project that uses LLMs.
The LLM is a dancing bear – it’s impressive because it does anything useful at all. It’s a wild beast: it’s highly unpredictable, difficult to train, dangerous if let loose, and expensive to feed, but it does attract a big crowd because it seems to be doing the impossible – sort-of – if you’re not very particular or demanding about the results.
The amusing part are the people who seem to believe that because this bear “dances”, soon every problem or task would be solved by ballets of trained bears.
This is a fantastic description of the zeitgeist, and I’m stealing it.
If anyone complains, I’ll claim I’m an AI, which apparently makes it ok.
Is this an advertisement for an AI service or a frontend for espionage disguised as an AI service?
Oh. You tagged one. And removed my comment. Now tag the rest of the LLM projects you all have posted over the last three weeks, assholes.