Agentic POC

Anyway, the backend is primarily Office 365 with PostgreSQL/PgVector for caching. Also used were components from Hugging Face, LangGraph, pgVector and exchangelib. Because the components are robust, and I have been doing this for a while, no AI coding assistant was needed. LangGraph was practically a point and click operation on its own.

One of the first challenges were hallucinations. That problem will continue to exist with questions that jail break the model, but a large chuck of the issues were mitigated by making sure the range of questions were supplemented with real data answers (fine tuning), and subsequently by using a mixture of LLM and traditional models to validate the content of the answers using embeddings / text review.

This project was completed in less than a day, and was able to answer a majority of the test questions with accurate responses. There is an escape hatch to a real person, and the incoming messages and responses are setup to be monitored by a real person to allow for continuous review. This allows humans to quickly reduce the noise, and work on the business at hand -- printing my photos on large canvas.

So where are the problems? The speed of the development. In a few hours, I was able to accomplish what once was impossible, and within the last 5 years would take months of development. This is hosted on an old gaming system running on an Nvidia 3090 graphics processor. Since it is a synchronous communication, the processing power is minimal, allowing 10 test users to see responses in under a minute while the system was also busy training an image processing model (more on this later).

My conclusion? If developers and technical founders are not using this capability to spend their time creating more solutions, then they are obsolete. But the real problem is, if everyone works together with compassion and humanity, the real opportunities are boundless, and if everyone is just working for a quick buck, then it seems we might be headed for another bubble.