Archimedes provides an AI tutoring platform for higher education that enforces professor-defined guardrails while giving instructors anonymized insight into class-wide confusion. The system integrates natively with Canvas, lets faculty upload syllabi and answer keys to configure the tutor, and surfaces ranked struggle points with pseudonymized conversation transcripts. Unlike general chatbots, it resists prompt injection and reveals what students don't understand without exposing individual identities.
Funding
Funding not disclosed
Founders
Product
Problem
Universities face a binary choice between banning AI tools entirely—leaving students unprepared for an AI-driven workforce—or allowing unrestricted access to general chatbots that hand over finished answers, bypassing the struggle that builds genuine understanding. Either approach leaves professors blind to what their students actually comprehend, forcing them to police browser tabs or accept academic integrity risks.
Solution
Archimedes provides an AI tutor that operates within each professor's explicit guardrails, turning every student interaction into an instructional asset rather than a liability. Faculty upload syllabi or assignment prompts, and the platform auto-suggests permission toggles—covering areas like brainstorming, grammar help, drafting, and direct problem-solving—which professors can refine with natural-language rules. The tutor is answer-key aware, using privately uploaded solution keys to nudge students toward correct approaches without ever revealing answers. Every conversation feeds an anonymized analytics dashboard that ranks genuine confusion points by follow-up weight and repeated attempts, with clickable pseudonymized transcripts showing the reasoning behind each struggle. A four-layer prompt architecture resists jailbreak attempts, and a built-in sandbox lets professors test their configuration against common bypass probes before deployment. The system lives directly inside Canvas as an LTI 1.3 tool, syncing rosters and enabling deep-linking to specific assignments without separate student accounts.
Target Audience
Primary customers are university professors, course instructors, and higher-education institutions using Canvas who want to integrate AI into their classrooms without sacrificing academic integrity or losing visibility into student understanding.
Features
- Four-layer prompt architecture that enforces professor-defined rules and resists prompt injection and jailbreak attempts
- Auto-suggested AI permission toggles parsed from uploaded syllabi, with natural-language custom rules and per-assignment overrides
- Confidential answer-key handling that provides right-track/not-quite feedback and hints without spoilers
- Anonymized confusion-ranking dashboard that weights genuine sticking points by follow-ups and repeated attempts, not raw question volume
- Pseudonymized conversation transcripts accessible by clicking any struggle point, revealing the "why" behind class-wide confusion
- Automated bypass probe that fires common jailbreak attempts at the configuration and scores each as held, partial, or leaked
- Native Canvas LTI 1.3 integration with roster sync, single sign-on, and deep-linking to specific assignments
- Sandbox chat mode for professors to test their own configuration as a student, with nothing saved