HokieLens

Inspiration

Every VT student registers for classes blind. The registration system tells you a schedule is valid, but not what it's actually like: can you make it from Torgersen to Whittemore in 10 minutes? Is that professor's grade history all over the place? We wanted students to see the real cost of a schedule before they commit.

What it does

HokieLens scores a set of courses out of 100 and breaks the risk down per course. It combines grade distributions, professor difficulty, class times, and building locations into:

a schedule risk score with per-course risk badges
walking times between back-to-back classes on a campus map, with tight walks flagged
an expected GPA range with a confidence label
what-if swaps: swap one course and watch the score move (49 → 19 in our demo)
Ask Gemini: plain-English explanations grounded in the schedule's own data

How we built it

FastAPI backend (Python) with a five-axis risk engine: workload, exams, travel, professor variance, grades. Grade stats use empirical Bayes shrinkage, and walking times use a Tobler hiking-function model
React + Vite + TypeScript + Tailwind frontend on Vercel, backend on Render
Gemini API for explanations, proxied server-side so the key never reaches the browser, with a grounded fallback if Gemini errors or times out
Live on hokielens.courses, with an agent card at /.well-known/agent-card.json (GoDaddy ANS)
Student impact analysis with a Databricks-importable notebook and CSVs (Deloitte x Databricks)
200+ backend tests, and fixture scores stay fixed so the model is deterministic

Challenges we ran into

Keeping the risk model explainable instead of a black box, making the scores stable and testable, and wiring generative AI in without letting it invent facts.