Early Notioc Testers: What We Learned from Student Feedback

## Early Notioc Testers: What We Learned from Student Feedback <div style="text-align: center; color: rgba(255, 255, 255, 0.7); margin-bottom: 2rem;"> <p style="margin: 0; font-style: italic;">A farewell retrospective from the Notioc team</p> </div> ### 1 — Setting the Stage When we started **Notioc** a couple months back, our mission was bold: give students an AI-powered teammate that could navigate their LMS as naturally as they do, surface the right context (lectures, PDFs, past homework) at the right time, and draft thoughtful, human-sounding answers—all without crossing academic-integrity lines. "Noti," our in-browser agent, was the most ambitious expression of that idea. Over the past semester, a hand-picked group of early testers put Noti through dozens of real-world quizzes, discussion boards, and reading assignments. Their candid feedback was the north star for every iteration that followed. Now, as we prepare to wind down Notioc and open-source the components that still show promise (most notably our **MCP server** that streams Canvas data via QUIC), we want to share what we learned—both the technical wins and the hard-earned lessons that ultimately led us to sunset the project. --- ### 2 — What We Tested | Sprint | Scenario | KPI we tracked | | ------- | ------------------------------------------ | --------------------------------------------- | | **S-1** | 3-question reading quiz (English 202) | *Time-to-first-draft* | | **S-2** | Weekly discussion prompt with peer replies | *Tone similarity* (Jensen-Shannon divergence) | | **S-3** | Mid-term study guide generation | *Citation accuracy* | | **S-4** | Locked-browser, open-note exam (Respondus) | *Agent stability* (crash-free minutes) | --- ### 3 — Five Things Our Student Testers Taught Us **Frictionless onboarding beats flashy features.** > "I spent more time pasting API tokens than actually trying the agent." > Replacing manual API keys with OAuth SSO cut drop-off by **41 %**. **Context > Parameters.** A 4-shot prompt with class-specific files beat a 32-shot generic prompt by **18 pp** on answer correctness—even when we used *smaller* models. **Authenticity is measurable.** Students wanted drafts that sounded *like them*, not a generic academic voice. Using a lightweight embedding to compare the agent's first pass with the student's prior submissions let us auto-dial verbosity and vocabulary in real time. **Privacy disclaimers must be *first-class UI*, not footnotes.** After we moved our data-usage summary to the top of the chat window, opt-in rates jumped from **64 % → 89 %**. **Price sensitivity is real—but tied to perceived *risk*, not dollars.** Most testers were fine with $175-200/month *if* we could guarantee zero plagiarism flags. Without that guarantee, even $15 felt expensive. --- ### 4 — How the Feedback Changed Notioc | Concern raised | What we shipped in response | | ------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- | | Token limits when ingesting large PDF sets | Incremental chunker + vector cache, reducing context window usage by **52 %** | | Agent "tone drift" mid-thread | Style-locking module that recalibrates every 8 turns | | Respondus lock-down kicking the agent | WebRTC overlay to mirror keystrokes locally instead of remote control, achieving **96 % pass-through** in mock Respondus sessions | --- ### 5 — Why We're Sunsetting Notioc 1. **Mission mismatch.** Over time, we realized our true excitement lies in **infrastructure for AI workflows** (e.g., the MCP server) more than in building a fully productized assistant for coursework. 2. **Academic-integrity gray zones.** Even with Integrity Radar, partnering with universities would never be a viable partnership strategy or a reliable business model. The constant policy flux would have forced us into full-time compliance work. 3. **Resource focus.** Maintaining Notioc's browser automation across six LMS variants and three proctoring tools meant chasing brittleness instead of advancing core tech. --- ### 6 — What's Next * **Open-sourcing MCP** Our QUIC-enabled Canvas scraper and orchestrator will be MIT-licensed on GitHub next month. We hope student devs run with it. * **Technical post-mortems** We'll publish deep dives on: 1. Managing long-context streaming with function-calling LLMs 2. Reinforcing user-specific writing styles via retrieval-augmented RLHF 3. Securing headless browser sessions under proctoring constraints * **Personal journeys** The founding team is joining hardware-AI focused startups & projects (where our skill sets shine) and will keep sharing lessons on our personal blogs. --- ### 7 — Thank You To every student who stress-tested notioc at 2 a.m., sent us crash logs, or bluntly told us when an answer "felt off"—*thank you*. Your feedback didn't just shape the product; it shaped our careers. Notioc is closing its doors, but the ideas—and the community—live on. See you in the next build. 🛠️🌓 — **Notioc team**