This blog category will focus on the upcoming release of the DigiMuse.AI project, which is the follow-on to the project that our team wrote for one of the nations leading resume providers. Tens of thousands of yearly revenue grew to over a million — all on our code. We had the resume process down. But I view career as a series of great journeys, of which the resume is the first step. What if we could build useful tools? Is AI a real thing for that? Those are the questions that DigiMuse.AI explores.
I have been fascinated with text analysis1 for decades, and now AI. I cannot seem to quit it. It fits what I like to code.
In the before time, I was working to improve Bayesian similarity, using bi-grams and tri-grams. Which seemed to improve analysis, and helped predict the next token. I hit a structural problem, words are a sparse information set, going from one token, to a pair, to trigram — made each “token” even sparser — yielding useless results.
Turns out, I was on the right track, just did not find the missing parts2. When I read “Attention Is All You Need“, I could not unread it. The Google team found the missing parts: embeddings, self-attention, and parameter sharing. I could not unsee the GPT diagram either, once I saw it.
I built my first LLM model in 2019 (Tensor flow of thousands of resumes), it was primitive, but cool. It was the first step of countless iterations and restarts. COVID, full time job, family, friends, and this thing is hard. There are lots of folks working on this, lots of offerings…
But I never lost site of DigiMuse.AI — maybe this solution will be easier, more effective, and there is only one way to find out. So I am reviewing my development logs and will be reposting them here on a regular basis. Timed to the release of the service, to allow others to share and help perfect the journey.
- Language analysis is the direct progenitor of modern AI, particularly todays large language models. ↩︎
- “I could have been a contender.” ↩︎