Headstart AI

The problem

Opening a Canvas assignment often means scrolling through dense instructions, a rubric, and a couple of attached PDFs before you can even answer 'what am I actually being asked to do, and how long will it take?' Students lose real study time just decoding the assignment before they get to the work itself.

The flow

Headstart is deliberately split into three pieces that mirror where a student actually is in their workflow, so the tool never asks them to leave it.

• Chrome extension — lives inside Canvas, detects the assignment page, and captures context (instructions, rubric, attachments) with one click, no copy-pasting.
• Web dashboard — where the guide streams in and the student can keep asking follow-up questions with the assignment context preserved, revisit past sessions, and plan study blocks on a calendar.
• Agent service — a FastAPI + LangChain pipeline on NVIDIA Llama-Nemotron that does the heavy lifting: native PDF text extraction with OCR fallback for scanned handouts, turned into a structured, streamed study guide via SSE.

Design decisions

• Streaming responses (SSE) instead of a loading spinner — a 30+ second generation feels immediate when the guide is visibly assembling itself.
• Chat stays assignment-aware across the whole session, so a follow-up question doesn't require re-explaining what assignment you're asking about.
• OCR fallback for attachments, since a meaningful share of professor-uploaded PDFs are scanned images rather than extractable text.

Research & validation

TODO: add findings from testing with students actively using Canvas this semester — where the extension's context-capture broke down, how accurate the generated guides felt against the real rubric, and whether the calendar planning view was actually used.

Outcome

TODO: replace with real usage metrics (guides generated, return usage across a semester, time-to-first-guide) once the current beta cohort has enough data.