Updates Aug 05
Astro photo theme
- css updates. before:

- after:

PGrid
- Wrote a scraper for passmark benchmark scores for each of the gpu chips.
- Using benchmark scores + specs, wrote a script that generates a prompt to generate a basic overview of each of the gpus and add a link to similar performance gpus in each gpu details page. Each of these descriptions were saved as a markdown file.

- Got those generated markdown files and displayed them for each gpu. works with SSG. happy with styling but font could be better. Advantage of prompt generated descriptions is that updated descriptions can be done for all gpus as prices change and new gpus come out.

- Started on ebay scraper. Ended up being more annoying than thought. they change around things like the format of the text of seller names and ratings and div class names.
- Settling on ebay scraper to work a bit different from the other scrapers. with ebay, i search for each gpu and then use ai to check if the listing title matches the gpu being searched. Have tried a couple of different models/prompts but hard to find 100% success rate with cheap models and i don’t want to manually verify as ebay listings dont seem as long lasting as typical ecommerce site listings.
Homeless help app
- New project to help a charity (Northwest Community Group) give more up to date resources to help people facing homelessness. Got some json from askizzy.org.au and converted to csv to then allow manual filtering and updating

Law statement of claim matrix app
- Started on a new project to use ai to create a pleadings matrix by combining text from multiple documents. Basic scripts up to turn scanned pdfs into images for processing by ai. Tried a few iterations to read the initial statement of claims document such that each claim is classified:
- convert document to json style markup based on paragraphs with claude. not 100% accurate
- convert document to markdown text then convert markdown text into xml style sections where each claim has its own tag and id. markdown conversion seems ok but tried a few different prompts with both gpt4-o and claude sonnet 3.5 but both haven’t been able to classify the claims properly.