Since launching on 'Show HN' 2.5 years ago: https://news.ycombinator.com/item?id=25678568 we have written 10s of thousands of resumes for free, all the while studying very carefully how expert human writers craft resumes and resume language for best success.
The biggest technical challenges have been:
- Sentence length. LLMs are trained on the corpus of all human writing and not a lot of that corpus is focused on writing sentences and lines of specific character lengths. As a result, getting the LLM to produce a sentence of, say, exactly 84 characters took a lot more effort than you’d imagine.
- PDFs. “A PDF file is often a combination of vector graphics, text, and bitmap graphics,” says Wikipedia. In many ways, PDFs are more of a drawing document than a writing document, and that often makes ingesting your past data impossible. The world’s text parsers are easily confused by PDFs, especially those using columns or non-standard text placement, which makes reliable parsing of your past experience difficult.
- Word choice. Modern American resumes are written in a particular idiom that combines upbeat achievement-oriented positivity with a terse and economical writing style to convey your capabilities and past successes. Hitting this consistently has meant training our internal writer (we call it TWAINN for Trained Writer AI Neural Network) on over a half-million sentences written by human experts. Our earliest efforts were often stymied by overfitting - for example, the one generation in which our model decided that the first bullet point of every work experience must begin with the word ‘spearheaded’. Fine-tuning is expensive and these blow-ups were frustrating.
I’ve always appreciated the support of the HN community, so please try it out and share your feedback!
p.s. Business model past two years has been tips. Since launching paid LinkedIn profiles and cover letter writing, that’s turned out to be enough to support the business going forward.