> SCRIPT RUNNING_
> GPU MAXED OUT_
:: OCR ON CHAOS :: LOCAL LLM, NO CLOUD
These 141 flyers are from my personal collection — physical originals from Calgary's underground rave scene, scanned by hand. The problem: rave flyers are intentionally unreadable. Warped fonts, neon on black, photocopied five generations deep, folded and creased. Standard OCR tools completely fail on them.
The solution was Meta's Llama 3.2 11B Vision model running fully locally on an RTX 3060 — the same vision-language model used to identify the Stadia game covers. Each flyer image was fed to the model with a structured prompt asking it to extract event name, date, time, venue, production company, and DJ lineup. The model handles the visual chaos that kills traditional OCR.
The raw responses were parsed by a Node.js script into a structured
flyers.json dataset, which powers the D3.js timeline and all the filter slicers.
The model was loaded with 4-bit NF4 quantisation — fitting 11 billion parameters
into 6.7 GB of VRAM on the 12 GB card.
Llama 3.2 11B Vision 4-bit NF4 quant BitsAndBytes HuggingFace Transformers RTX 3060 12 GB Python Node.js D3.js
:: COMMENTS ::