- Art portfolio anonymizer for a company I was doing consulting for. Scrubs real portolios to protect candidate privacy and portfolio/art poaching when jobs are outsourced to external recruiters. Approximates art, photography, and design without it being a direct copy of the original work.
- Car Talk guys. Had a car problem but figured the voluminous amount of information from Car Talk would probably come up with a better solution. Transcribed the library of episodes using speaker identification WhisperX (Click/Clack) and ran them through an OpenAI-“like” API using free embeddings models, langchain, chromadb, and off the shelf code to answer either individually or as a team. This was to simulate a conversation between the two of them and the “caller”.
- Stobe. Similiar idea to above but built around a YouTube celeb that used to hop trains and used to do a lot of talking about life while waiting for them (it’s like 90% wait). I wanted to capture all of these conversations to build up a virtual personality for him after he passed away. I thought it a little creepy but at the same time sort of heartwarming. BONUS: you can chat with Stobe and find the best trains to hop going anywhere in the U.S. and which ones to avoid! Apparently BNSF!
- Fine tuned small LLM’s and/or built LORA’s to be good at sorting and tagging data for building datasets faster. E.G. iterating through images using a lightweight Vision edge model (e.g. moondream2) to find low quality or blurry images, tagging and sorting them off, resizing, trashing, etc. Also did this for sorting through Flickr for actual night-time images (CC license of course) to create LORA’s. These LORA’s in turn would supplement models that are generally not great at night-time scenes (SDXL mostly although FLUX has that problem sometimes). This is turn was for generating imagery for the children’s book that is taking forever and a day to get organized. But also for making really cool realistic night photographs.
- Workflows, workflows, workflows. For visual stuff I think one should go no further than ComfyUI for testing and learning AI/ML. Production, no. It’s like Visio+++++. So built hundreds of workflows, really scripts, that do various things in very specific ways (for me). If I have the time I can spend 24 hours straight just messing with ComfyUI. Addictive.
So yeah, a lot of AI for 2024.