Cadent — AI Study Planner for College

The problem

What I built

• Built a macOS desktop app that connects to a student's Canvas, reads their real coursework — assignments, syllabi, announcements, course pages, and PDFs — and turns it into a daily study plan, entirely on-device.
• Designed a two-phase planning engine that separates understanding the work (extracting each assignment and its true deadline) from scheduling it (breaking an assignment into a few real steps and placing them on days, capacity-aware, so a small assignment lands as one focused block the day before it's due rather than smeared across the week).
• Built a grounded extraction layer that dates each assignment only from the claim that actually states its deadline — so a reminder mentioning three due dates can't cross-wire one onto another — and a source-authority model for when a professor's announcement should override a stale Canvas field.
• Added a conversational layer: a student can say "I'm busy tonight, reschedule everything" or move a single assignment to another day, and Cadent replans — warning first, and letting them confirm, when a move would land past a deadline or overload a day.
• Built a verification ("guard") layer that reviews every proposed change before it's committed, so the planner can never silently accept lateness or drop work, and a local encrypted store so a student's entire academic profile stays on their own machine.
• Shipped the gateway that serves the language-model reasoning, with tiered models chosen per task to hold measured cost to a few dollars per student per month — and made the whole thing invite-gated, because the invite count is the spend ceiling.

What I had to learn

• Grounding an AI system against a source of truth. The hard part of "read the syllabus" isn't reading — it's refusing to invent. Most of the engineering is machinery that forces every deadline, task, and claim back to the exact words a document actually wrote, and treats a missing fact as missing rather than guessing.
• Designing for calm, not features. A study planner that overwhelms a stressed student has failed regardless of correctness. Decisions like "one focused block, not a scattered week" and "warn-then-confirm instead of hard-refuse" were product-psychology calls that shaped the architecture underneath them.
• Where to put trust in an AI pipeline. Deadlines and coverage should come from structured provider data whenever it exists; prose should refine, never fabricate. Getting that hierarchy right — and building the reconciliation that decides when several sources describe one assignment — was the difference between a plan a student can rely on and one they have to double-check.
• Building a correctness-critical AI product solo, end to end: ingestion, extraction, scheduling, a guard layer, encrypted local storage, the desktop client, and the model gateway all have to agree on one model of what an "assignment" is.

Results

• Working private beta: signed macOS builds, live model gateway, real students onboarding from their own Canvas.
• Measured cost held to a few dollars per active student per month through tiered model routing.
• Presented to VCs and mentors who have responded positively to the idea and are advising its direction.

Stack