Eduvision Ai: An offline, syllabus-driven system for generating narrated animated video lessons In computer science

Eduvision AI model demonstrates that a locally run, freely available combination of a language model, a text-to-speech model, a speech-alignment model and conventional rendering code can generate syllabus-specific, narrated, animated Computer Science lessons without any paid service, user account, or internet dependency. The central finding of the project is architectural rather than purely technical: reliability is achieved not by prompting the language model more carefully, but by restricting the model to the single task it performs well — writing prose — and delegating every task that requires exact counting, arithmetic, or structured factual  accuracy  to  deterministic  code  that  is output is trusted. This division of responsibility allowed the system to reach 808 passing automated tests, syllabus coverage of 286 topics, and fully offline operation at zero recurring cost, making it a viable tool for a student, a lecturer without a teaching-aid budget, or an institution with unreliable internet connectivity. Future work identified during development includes extending the native diagram library with additional plot and matrix templates, which is expected to raise the proportion of topics receiving a fully code-derived animation from 82 to approximately 93 of the 286 syllabus topics, together with a formal human evaluation of learning outcomes compared with conventional revision materials.

[1] S. K. D’Mello and A. Graesser, “AutoTutor and Affective AutoTutor: Learning by Talking with Cognitively and Emotionally Intelligent Computers,” ACM Transactions on Interactive Intelligent Systems, vol. 2, no. 4, pp. 1–39, 2012. [2] J. Chen, Y. Liu and X. Wang, “Automatic Video Generation from Text Using Deep Learning Techniques,” IEEE Access, vol. 8, pp. 123456–123467, 2020. [3] T. Brown et al., “Language Models are Few-Shot Learners,” in Proc. NeurIPS, 2020. [4] A. Vaswani et al., “Attention is All You Need,” in Proc. NeurIPS, 2017. [5] K. He et al., “Deep Residual Learning for Image Recognition,” in Proc. CVPR, 2016. [6] A. Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” in Proc. ICML, 2021. [7] Y. Wang et al., “Tacotron 2: Natural TTS Synthesis,” in Proc. ICASSP, 2018. [8] R. Rombach et al., “High-Resolution Image Synthesis with Latent Diffusion Models,” in Proc. CVPR, 2022.