Arcanum AI — Flagship Educational Platform
International and non-native English speaking students face steep cognitive hurdles during fast-paced academic lectures filled with dense jargon, lack of contextual scaffolding, and absent live feedback loops for educators.
My Role & Contribution
Co-Founder & AI/Backend Lead: Designed the core concept simplification pipelines, prompt orchestration logic, and high-concurrency FastAPI microservice. Built the prototype in a 36-hour sprint at TetraTHON 2026 (6th place / 160 teams) and currently directing the transition to multi-agent architecture.
Proposed Solution
Engineered Arcanum AI as a dual-facing system: students receive adaptive, on-demand conceptual simplifications and multilingual term breakdowns, while educators view aggregated, anonymized comprehension heatmaps. As AI & Backend Lead, I engineered the asynchronous inference routing pipelines, prompt engineering logic, and backend telemetry handlers.
- Contextual concept simplification adapting complex academic definitions into digestible explanations
- Multilingual translation pathways supporting non-native English speakers in real-time
- Classroom intelligence telemetry tracking student comprehension bottlenecks for educators
- Asynchronous FastAPI backend engineered for low-latency query processing
- Secured 6th Rank among 160 teams at TetraTHON 2026 (Navrachana University & ISEN Méditerranée)
Architecture & Dataflow
Implementation Status & Roadmap
Implemented
- Real-time contextual academic concept simplification engine
- Multilingual translation layer for lecture terms
- Student comprehension feedback telemetry dashboard
- High-concurrency async FastAPI REST gateway
- Deployed live web prototype at tetra027.web.app
Planned Roadmap
- Multi-agent orchestration (Tutor Agent, Terminology Agent, and Evaluator Agent)
- Syllabus-based Retrieval-Augmented Generation (RAG) referencing specific course textbooks
- Speech recognition pipeline generating aligned, searchable lecture transcripts
- Adaptive spaced-repetition quiz generator calibrated to each student's weak points
Engineering Trade-offs & Limitations
Rationale:During the initial 36-hour sprint, we prioritized lightweight streaming responses and prompt-engineered contextual layers over offline model fine-tuning to guarantee reliable sub-second response times during live jury evaluations.
Results & Evaluation
Won 6th Rank out of 160 competing teams at TetraTHON 2026 (organized in partnership with ISEN Méditerranée, France). Deployed live at tetra027.web.app and actively expanding into multi-agent syllabus RAG.