Back in May of 2019 my LLC partner and I started exploring how artificial intelligence could impact education. Unfortunately tools for building AI driven applications were almost nonexistent at the time [this was two years after the Attention is All You Need paper, Large Language Models had yet to become commercially available].
Fast forward to February of 2024. By then OpenAI's GPT3 models had been released; LLM inference hosting providers were emerging with well documented APIs, and open source models were meeting or exceeding commercial models. As the R&D lead at Last Myle this was the breakthrough we needed to start building more complex AI experiments.
At the same time I was eager to leave LinkedIn; it was a legacy professional network model that didn't fit how I wanted to express my career and passions.
So with LLM inference APIs and multi-model routers, and modern web frameworks I was able to start building and running experiments for a new personal site experience.
The first experiment was focused on evaluating different models. Search for Intelligence was a single evaluation workflow to score text and image responses. I also built a rudimentary memory system to store chats and evaluation scores.
The next experiment applied scoring analysis of models from Search for Intelligence, to build a personal RAG pipeline. This enabled people to chat with a hybrid of my curriculum vitae. It was also an excuse to test Google's NotebookLM and the quality of podcast output.
From there I started experimenting with agents; first with Microsoft Research's AutoGen, then CrewAI, then Vercel's AI SDK. Simultaneously I started experimenting with Terminal UIs running in the browser; then building a hosted service to test multi-tenant performance. The summation of these experiments lead to my agents terminal system which has now been running continuously for weeks.
My agents terminal system has multiple agents. The one I'm most excited about is the matching agent that converses with you to find our collaborative match. We'll see who gets paired with me; and what we build together.
Back at this site; it's also been updated to experiment with primitive agentic system visualization. If these become popular then I might pipeline them into sandbox environments; and leverage what I built for sandbox hosting.