From VHS to Vision Models

I had the opportunity to contribute another post for the CommBank Technology Medium blog. This time around on something a little less speculative: Preserving CommBank’s Archives with AI: From VHS to Vision Models. Amidst the recent spate of AI-powered engineering initiatives, it felt nice to work on something that felt ethically aligned with my values. There’s a pretty severe AI-instigated dislocation happening, and I think we’re only just starting to understand and feel the effects of it. But recorded audio transcription and image description for digital preservation ticks the boxes for me. The work simply wouldn’t have happened without a cost-effective capability. On the technology side, I’m also a big advocate of open-weight models, and it’s pretty amazing what even (small!) transformer models running on commodity GPUs can accomplish now. As an added bonus, I got to play around with one of my favourite techniques: structured outputs/constrained decoding. What a nifty application of…