I’ve been building this for a while and it finally works end to end. It listens to your system audio, transcribes it live with whisper.cpp, and when you hit a hotkey it answers questions using your CV and the job description. Answers and embeddings both run through Ollama, so by default nothing leaves your machine.
The Ollama side is the part I put the most work into:
- A setup wizard reads your GPU and recommends a model that fits in VRAM alongside the speech model, then pulls it for you.
- You can find, download, switch and delete models from inside the app. Before a pull it checks the tag exists and that you’ve got the disk space, counting layers you already have.
- It won’t let you delete the model that’s currently answering or the one doing embeddings.
- When you start an interview it warms the model up, so the first answer doesn’t wait for it to load. Keep alive is adjustable in settings.
- If the model supports thinking you get a reasoning toggle, off by default because the thinking takes too long in a live call.
- RAG over your CV and the job description uses nomic-embed-text, also through Ollama.
Whisper runs on the Vulkan backend, so it should work on GPUs from AMD, Intel and NVIDIA. Cloud keys for Anthropic, OpenAI, Gemini and Groq are there if you want them, but they’re optional.
This is my first release and I’ve only properly tested it on my laptop with an RTX 3070, so it’d be a big help if you gave it a go on yours, especially on AMD or Intel cards and desktops other than KDE. The .deb is for Ubuntu 24.04 and newer. Fedora, Arch and openSUSE need the source install in the README. Ollama isn’t bundled, but the wizard shows you the install command if it’s missing. If anything breaks, please let me know in the issues :)
Also, which models are you finding best for quick, short answers on 8 GB and 16 GB cards? The wizard’s picks are a starting point and I’d like to tune them. Thank you <3
https://github.com/PSYEONE/interviewr-ai
How should a person implement a stt prompter for use in remote interviews without using llms?
I’m genuinely interested in hearing the alternative.
Um…
By just not doing so and answering genuinely?
Not what I asked but let’s explore this:
No thank you, I will instead use every tool at my disposal to get a job I want. If a software package can more reliably alleviate the brain farts and social discomfort associated with zoom meetings I’ll use it over taping a list of reminders or a mind map to the edge of my monitor any day.
If I’m interviewing you and get a BUFFERING blank stare response while your AI generates a response to the simple screening questions I ask, then you’re not getting hired. Authenticity in an interview is way more impressive and valued than a “perfect” studio answer.
Also, we can see your eyes scanning your generated script when you respond. You won’t fool anyone that is worth working for.
Edit: spelling error
And even if you did somehow manage to, once you’re actually doing the job, you’d get found out pretty quickly… I’ve heard of people trying this nonsense, and there is no scenario where it works put for anyone, it wastes everyone’s time, and just puts up more barriers to stop genuine hard working people accessing the jobs that actually fit them. Yes, the interview landscape can be brutal at the moment, but ‘tools’ like this are the problem, not the solution.
I’d say congrats on the ratio, but this one was a piece o’ cake lol
Who’s the artist of these anti-slop images?
I’m sorry, but this is one of those “trying to fix a problem that does not exist” programs that serves no purpose other than clog the linux env. unnecessarily. I mean, it even requires vulkan? Seriously?
And no. This is not a “AI bad” post.
I’d go one step further. This kind of ‘tool’ is not fixing a problem that doesn’t exist. It is actively causing a problem that didn’t exist, and doesn’t need to (along with the charlatans peddling the AI application-slop ‘services’). I’ve heard from hiring managers of people trying things like this, and it:
- Categorically does not work, no-one is fooled
- Results in extra barriers needing to be erected.
Ultimately, it makes an already brutal interview landscape 1000x worse, for no benefit to anyone.
Fair enough if it’s not for youu. On the Vulkan bit though, it’s only there so whisper can run on the GPU, which is what makes live transcription fast enough to keep up with a call. If there’s no GPU it falls back to the CPU by itself. The only thing it actually pulls in is libvulkan1, which is a small loader that’s already on almost every Ubuntu desktop :)
Don’t listen to that person. His last post was about his Winamp clone that only plays oggs made this year.
There is no person who should say that kind of stuff less than him.
Thanks, appreciate it. Part of why I built it is that a lot of first round interviews now are run by an AI with nobody on the other end, so it felt fair to have something on my side too. It works for normal calls and meetings as well, but that was the push :)
I work in data and the issue here isn’t AI. Nobody in data engineering cares enough about AI to make anti-AI memes like I’ve seen in comments.
The problem here is this tool is only viable for the non-technical aspect of the interview process. In your repo’s gif image you mention you have PostgreSQL experience. Great! Show me. We aren’t going to take your word for it, you’re going to be asked to whiteboard queries based on a test setup we provide in a live environment that we provide.
Even if your AI agent is locally running, we are going to be watching it in real time. Even if your AI agent is running, we are going to be looking for things you likely didn’t include in the prompt. Did you ask it to write the query or write the query in an optimized manner? Those are two different things (and a very important distinction when working with large data sets).
Specific to AI: We don’t have any testing specific to how you use AI agents, but I have seen dev groups which have dedicated environments they build with throw-away API keys that will test how you use AI in code authoring. So that is something to consider depending on what role you’d be interviewing for.
In the end, at least on the technical side, we don’t really care what your CV says. We’re going to put you through the same test process regardless.
EDIT: After reading one comment, I want to say AI interviewers are a cancer. They act as negative modifiers to throw out candidates based on arbitrary keywords. Personally, if I encountered one I would seriously consider cancelling (obviously dependent on my options). If HR or other depts are so overworked or so lazy that they cannot actually meet me, then that is a huge red flag about the company.
The point of a tool like this is the non-technical part of the interview process.
It’s not a problem that it only applies to that component.
That’s like saying the problem with a Phillips head screwdriver is that it doesn’t turn torx head bolts. Of course it doesn’t. It’s not meant to do that.
My Phillips driver isn’t deficient because it can’t remove the torx bolts that hold my nvme tray to the frame of my laptop, its job is to get me past the Phillips head screws securing the bottom cover in place so I can access those Torx bolts with the correct tool.
Thanks for the detailed reply, that’s fair. It’s really aimed at the screening and behavioural rounds, especially the ones run by an AI, not a live technical test. Agree with you on AI interviewers too, that’s what pushed me to build it. Also the CV in the demo is just a sample one, not mine :)
