NEXUS LEARN AI

Inspiration

Students study from lecture videos, textbooks and slides that are scattered across different places. When they ask a general chatbot for help, two things go wrong: the answers aren't tied to their course, and the chatbot has no idea what they already understand.

We wanted a tutor that is trustworthy about the subject and aware of the learner. That idea became NEXUS LEARN AI: an AI tutor that knows your course and knows you.

What it does

NEXUS turns course material into a source-linked Course Universe and tracks each student in a Learning Twin.

• Grounded tutor: answers are extracted from the course material and cited with their exact location (page, slide or timestamp). A Grounding Shield labels each response as grounded, partially supported or not covered, and off-topic questions are refused instead of guessed.
• Adaptive quizzes: questions are verified before they are shown (the answer must appear in the cited source, the options must be distinct, and the question must not repeat). Difficulty rises after a correct answer and falls after a wrong one.
• Learner model: mastery is tracked per topic with a transparent rule, $m' = (1-\alpha)\,m + \alpha\,e$, where $e$ is the evidence from the latest answer, weighted by difficulty. Every change is logged, so each before/after number comes from recorded events.
• Misconception detection and Next Best Action: wrong answers are analysed by what the student chose, and rules recommend what to study next.
• Also included: a diagnostic for new students, Mastery Battle, Teach-It-Back, a study planner, a course flow map and an Evaluation Center.

How we built it

We built it in two layers.

1. Working prototype: a single-page web app that runs in the browser with no API keys. Retrieval uses IDF-weighted term matching, and the Grounding Shield treats at least 75% query coverage as grounded, at least 40% as partial, and anything lower as not covered. We tested the real script by rendering every page and running complete quiz flows.
2. Full-stack architecture: a FastAPI backend with SQLAlchemy and ChromaDB, parsers for PDF (PyMuPDF), slides (python-pptx) and video (faster-whisper), and a React, Vite and Tailwind frontend. The logic modules (mastery model, recommender, chunker, citation validation, answer verification) have unit tests.

One design rule runs through it: each content unit stores its source location once, and citations only reference those units. A citation can never point to something that wasn't ingested.

Challenges we ran into

• Honest grounding: making the system say "not covered" reliably, without refusing questions it should answer. We had to tune word matching (for example, "classify" vs "classifies").
• Question quality: a wrong answer key corrupts the learner model, so we added verification checks and a no-repeat guarantee.
• Honest evaluation: in our simulated-student test, adaptive selection removed repeated questions (0% vs about 31% for random selection) but did not beat random selection on weak-topic gain. We report that result as it came out.
• Scope: we chose one complete, working learning loop over a wide but shallow feature list.

Accomplishments that we're proud of

• A complete loop (ask, assess, diagnose, adapt, reassess) where every number shown is computed from logged events.
• A tutor that cites real sources and refuses what the material doesn't cover.
• On our team-built set of 14 questions (10 answerable, 4 off-material), the prototype reached 100% hit@1 and hit@3, refused all 4 off-material questions, and wrongly refused none of the answerable ones. This is a small test set, so we treat it as a sanity check, not proof of general accuracy.

What we learned

• Grounding is mostly an engineering problem. Thresholds, citation validation and refusal behaviour matter more than which model you use.
• A transparent learner model is easier to test, debug and explain than a black box.
• Reporting weak results honestly makes the rest of the results more credible.

What's next for NEXUS LEARN AI

These are not implemented yet:

• Connecting the PDF, slide and video parsers end to end, so real uploads get cited answers.
• Image and diagram understanding with a vision model.
• Evaluation with RAGAS (faithfulness, answer relevancy, context precision and recall).
• Short-answer and numerical questions, plus cross-model answer verification.
• Flashcards for weak topics, a forgetting-curve study schedule, and Hindi or Tamil explanations with audio.

Built With

• claude
• css3
• fastapi
• github
• html5
• javascript
• python
• react
• sqlite
• svg
• tailwindcss
• typescript