AI Team & Prototypes

Starting with research

In summer 2023, Jukabo started AImpacts to find practical uses for language models in business intelligence and internal tools. The first stretch of the project was mostly research. While we looked for four people to join the team, I tried to understand the technology well enough to narrow down what we should focus on.

I spent that time digging into transformers, fine-tuning, quantisation, embeddings, and vector databases. I wrote up the fundamentals and open questions in shared documents. Those notes gave us a common starting point once the team came together.

Coordinating the team

When four colleagues joined me, my role changed. I set the broad direction and planned what we would investigate. Day to day, everyone worked fairly independently for a few weeks at a time, then brought their findings back to the group.

We used prototypes to answer the open questions, building just enough to test an idea with real data or on an internal task. My colleagues used that approach to build projects such as Qortex, while I worked on the three prototypes covered below.

The three prototypes I built

Cognos AI assistant

I built a custom component that added a chat assistant to IBM Cognos reports. It could send report data to OpenAI and answer questions about what was on screen. We presented the prototype at a Cognos User Group meeting. The implementation has its own project write-up.

Meeting Utilizer and RedSchnell

We were having a lot of sales calls, and finding a particular detail afterwards was awkward. I built the Meeting Utilizer to transcribe uploaded recordings and let users ask questions about the transcript. For recordings that did not fit into the model’s context window, I added a rolling summary. I later rebuilt the tool in Next.js under the name RedSchnell to improve the interface and stability.

Jobmarket Analyzer

We also wanted a clearer view of the freelance market. The Jobmarket Analyzer collected listings from relevant job boards, converted them into embeddings, and stored them in a vector database. A React frontend showed changes in the market and helped us look for potential clients.

Smaller experiments

Alongside the three applications, I kept exploring the underlying technology. I built a fine-tuning pipeline for Llama 2 in Google Colab, using LoRA and PEFT to work around the limited hardware. The point was to see what adapting a model to company data actually involved.

I also built a small semantic-search demo for sales conversations. It let people try an embedding-based search instead of sitting through an explanation of vector databases.

Result

By April 2024, I had built three larger prototypes and several smaller technical experiments while coordinating the team’s wider work. My colleagues had built additional prototypes, and together we’d documented what we learned along the way.

Looking back, I’m proud of what we built, and I really enjoyed working with that team. Much of what I learned about machine learning and the AI tools and infrastructure we used still informs my work today.