Last updated: 22 October 2026. Written by Ritesh Kanjee, director of Augmented AI.
I sold a retrieval-augmented generation (RAG) agent for $4,500 to a university that wanted an AI tutor inside its learning management system. They needed predictable costs and no external API fees, so the agent runs entirely locally with offline models on an on-site GPU. Students get answers from their course material without relying on cloud models.
A university client approached me with a unique challenge. They run a learning management system (LMS) and wanted a Retrieval-Augmented Generation (RAG) solution to create a better learning experience for their students.
But universities operate under a unique set of constraints.
What problem did the client have?
Using API keys or subscriptions isn’t simple in a large institution. The process involves someone putting costs on a personal credit card, followed by a tedious reimbursement process. Worse, it’s nearly impossible to predict monthly usage and costs when hundreds of students are using a chatbot. They needed a solution free from recurring fees and red tape.
Their goal was to build an internal learning tutor, similar to Google’s NotebookLM. The concept was straightforward: students could select specific course documents, then chat with an AI to understand the material better or generate quizzes on the topic. The key was that they didn’t want to load all documents at once; the system had to be selective.
To avoid unpredictable costs, they decided to run everything locally. This meant using offline models for both the embeddings and the chatbot itself, all powered by a dedicated on-site GPU. The entire system had to be self-contained, with no reliance on external databases or cloud models.
In short, the requirements were:
- No recurring fees, card reimbursements or unpredictable monthly usage costs.
- A tutor similar to NotebookLM, where students pick specific course documents, then chat or generate quizzes.
- Everything local: offline models for embeddings and the chatbot, on a dedicated on-site GPU, with no external databases or cloud models.
What was the technical hurdle?
This project was more complex than deploying a standard chatbot. The central challenge was a technical one. Could we individually vectorize and embed each document beforehand and then allow students to select from them on the fly? Or were we limited to a single, massive database of all documents?
The alternative, embedding documents in real-time, was not an option. The latency would destroy the user experience. We had to find a way to make the system both flexible and fast.
After some research and development, we found our answer. It was possible to pre-process and embed each document separately, storing them for instant retrieval. This was the breakthrough the client needed.
What did I build?
We built a containerized chatbot dashboard that plugged directly into their LMS. It delivered the speed and flexibility they required, allowing students to select specific sources for their study sessions without any lag.
This project is a perfect example of moving from Level 1 automation chaos to a structured Level 2 system. It solved a specific, high-value problem and created a repeatable asset. This single workflow is one of several that have helped me build a business generating over $10,000 per month.
If you want access to the high-impact workflows I’ve built for clients, you can find them in our Corporate Automation Library. We have a full suite of automations you can begin using right now.
We have over 50+ high-impact, high-ROI automations, with 2 to 4 new corporate automations uploaded weekly.
