Catch cheaters with AI. Local processing, zero data collection, full privacy.
Drop in gameplay footage, train a model on your clips, and let it figure out who's being a bit too accurate.
Feeds your gameplay through a Vision Transformer, extracting 16 frames per clip. Supports most formats - MP4, AVI, MOV, MKV.
Uses librosa to detect kill sounds and automatically snips 2-second clips around each one. Pretty neat actually.
Label clips as cheating or legit, fine-tune a ViT model on them. The more data you feed it, the better it gets.
Get scores from 0.0 (definitely clean) to 1.0 (yeah that's sus). Frame-by-frame breakdown included.
Everything stays on your machine. No cloud uploads, no tracking, no sketchy data collection. Just you and your GPU.
Save everything as JSON for later. Build reports, track patterns, or just keep records for yourself.
Got tired of getting beamed by players with aim that was just a bit too perfect, you know? Started wondering if AI could pick up on the same patterns I was noticing.
Turns out Vision Transformers are pretty good at spotting unnatural aim patterns. This project is the result of that curiosity - a local tool that analyzes gameplay and gives you confidence scores.
It's not perfect, and it's definitely not Riot's Vanguard, but it's a fun experiment in applying ML to gaming problems.
Should take about 5 minutes. You'll need Python and ideally a CUDA GPU.
# clone the repo
git clone https://gitlab.com/Vyapari-Dev/vanguard.git
cd vanguard
# setup virtual environment
python -m venv venv
venv\Scripts\activate
# install dependencies (first run downloads models
~2-3GB)
pip install -r requirements.txt
# run the app
python main.py
Pretty straightforward once you get it running.
Load up your Valorant gameplay. The app will analyze the audio track and auto-extract moments where kills happen.
Sort clips into folders - cheating/ for sus plays, legitimate/ for normal gameplay. More clips = better model.
Hit the train button, grab a coffee. 10-20 epochs usually does the trick. Model saves automatically.
Feed it new gameplay and get confidence scores. Anything above 0.7 is worth a closer look.