2026Fine-TuningOpen-SourceCompiler Messages
Fine-Tuning Open-Source Models as a Viable Alternative to Proprietary LLMs for Explaining Compiler Messages
Tests open-source models against proprietary ones for keeping student data in-house.
Forthcoming
Research
Building AI with educational instincts
Years of peer-reviewed research into fine-tuned language models that explain, guide, and teach, not just answer. This research powers Juno’s tutoring engine today.
Publications
7 papersPeer-reviewed and preprint research from our team, and what each paper contributes to Juno.
arXiv 2025Fine-TuningOpen-Source LLMsPedagogy
Narrowing the Gap: Supervised Fine-Tuning of Open-Source LLMs as a Viable Alternative to Proprietary Models for Pedagogical Tools
The technical basis for Juno’s privacy-first, self-hosted approach.
arXiv preprint arXiv:2507.05305
Read paperarXiv 2025Fine-TuningPedagogical AgentsProgramming Education
Supervised Fine-Tuning LLMs to Behave as Pedagogical Agents in Programming Education
The method behind Juno’s “guide, don’t give away the answer” behaviour.
arXiv preprint arXiv:2502.20527
Read paperSIGCSE 2025Compiler IntegrationConversational AIDebugging
Compiler-Integrated, Conversational AI for Debugging CS1 Programs
The system now running as Juno’s tutoring layer.
Proceedings of the 56th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2025)
Presents a compiler-integrated conversational AI system that helps CS1 students debug their programs through pedagogically-grounded dialogue.
arXiv 2024Supervised Fine-TuningPedagogical LLMs
Towards Pedagogical LLMs with Supervised Fine Tuning for Computing Education
Early groundwork: fine-tuning, not prompting, makes a model teach.
arXiv preprint arXiv:2411.01765
Read paperKoli Calling 2024Fine-TuningError Explanations
Fine-Tuning Large Language Models for Better Programming Error Explanations
Direct predecessor to Juno’s error-explanation feature.
Proceedings of the 24th Koli Calling International Conference on Computing Education Research
Explores fine-tuning large language models to produce better, more pedagogically appropriate explanations for programming errors encountered by novice programmers.
SIGCSE 2024DCCCompiler ErrorsLLMs
dcc --help: Transforming the Role of the Compiler by Generating Context-Aware Error Explanations with Large Language Models
The origin point. Still running at UNSW with 50M+ uses.
Proceedings of the 55th ACM Technical Symposium on Computer Science Education V. 1 (SIGCSE 2024)
Introduces dcc --help, a system that transforms the compiler into a pedagogical tool by using LLMs to generate context-aware explanations for error messages.