Final Year Project · AI / EdTech
APIA
An AI tutor that guides first-year students using the Socratic method: instead of giving the answer, it asks the right questions to help them solve it themselves.
Built on a fork of Open WebUI, with a local deployment for the school and a RunPod integration for testing larger models on rented GPUs.
The Problem
When students copy-paste code from ChatGPT instead of working through the problem, they skip the part where actual learning happens. They get the assignment done but never build the intuition.
"We found that using AI assistance led to a statistically significant decrease in mastery. [...] Participants in the AI group scored 17% lower than those who coded by hand."
The Solution
APIA acts as a Socratic tutor. It doesn't give answers. It guides students with questions and hints, forcing them to think critically while still getting 24/7 personalized support.
Our Work
I was the Technical Lead. Our work focused on benchmarking open-weight language models, vector databases, OCR engines, and embedding models to choose the components for the tutor. We selected Qwen-Coder for the language model and built the RAG pipeline around the school's course materials.
Starting from Open WebUI
We forked Open WebUI so we could reuse its chat interface and user management. That saved us from rebuilding those parts and gave us a base to adapt for the school.
Our focus was on the AI components: comparing models, document processing, retrieval, and embeddings, then putting them together into a tutor that uses the course content to guide students.
How it works
A RAG pipeline that reads official course PDFs, so the LLM can reference specific class concepts when guiding students.
Fast Inference
vLLM with 4-bit quantization. One server handles multiple students at the same time with minimal latency.
Local deployment
The school deployment runs the tutor and its models on campus hardware. Remote GPU pods provide another option for testing larger models.
Context-Aware
Docling OCR reads actual course materials. The AI references the exact exercises students are working on.
Testing larger models on RunPod
We also implemented a feature to rent a remote GPU pod on RunPod and connect it to APIA. The pod runs Ollama or vLLM to serve the model.
This lets us test heavier open-weight models on rented hardware and use the results to inform our model selection. The remote testing option sits alongside the local school deployment.
Results
Presented in February 2026. Selected by the school for a pilot program in 2026-2027.
2026–2027 pilot selection
Selected for a pilot in introductory programming labs during the 2026–2027 academic year.
Self-contained deployment
Packaged so the school IT team can install it without us