Empowering Mobile-Only App Generation — Offline AI Code Generation with App Inventor

As digital tools become more accessible, creating software is becoming a powerful way for anyone to make real-world impact. Computational action—the idea that learners can build computing artifacts with authentic relevance to their lives and communities—reframes computing as a tool for empowerment. Low-code platforms like MIT App Inventor support this vision by fostering digital agency through purposeful creation. Recent advances in large language models (LLMs) expand these possibilities further by enabling code generation from natural language, offering a timely opportunity to lower the barrier to app creation. MIT App Inventor has long championed accessibility, allowing even young learners in underserved regions to build meaningful mobile apps. Its natural language tool, Aptly, enables users to describe app ideas and generate functional code. However, Aptly’s reliance on cloud-based LLMs limits access for users without stable internet—often those who could benefit most. This thesis addresses that challenge by enabling AI-powered app creation to run entirely offline on mobile devices. We fine-tune and quantize LLaMA 3B using QLoRA and deploy it on iOS with MLC LLM, enabling on-device inference without internet. We also introduce a custom evaluation framework tailored to Aptly’s grammar, combining a Tree-sitter parser and a modified CodeBLEU metric to assess both semantic and syntactic quality. Using curated evaluation datasets, we benchmark out-of-box and fine-tuned models across prompting strategies. In our evaluations, fine-tuned GPT-4.1 achieved the highest normalized CodeBLEU score (0.36 ± 0.12) and parsed over 81% of completions, outperforming its baseline by more than 5%. QLoRA-finetuned LLaMA improved parseability by 11.7% over its base model, showing progress in adapting smaller models to the Aptly domain, though semantic fidelity remains a challenge. Our results show that offline natural language–to–app generation is feasible, and that smaller models can be adapted to the Aptly domain. By lowering the technical and infrastructural barriers to app creation, this work lays the foundation to empower AI-assisted programming that is accessible, offline, and on the phone.

Name thesis:degree_name

Master

Department dc:contributor.department

Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science

Grantor dc:publisher

Massachusetts Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Yuan, Joyce

Advisor dc:contributor.advisor

• Abelson, Harold

Statement dc:rights

• In Copyright - Educational Use Permitted
• Copyright retained by author(s)

Licence dc:rights.uri

• https://rightsstatements.org/page/InC-EDU/1.0/

Handle dc:identifier.uri

https://hdl.handle.net/1721.1/163028

OAI identifier oai:identifier

oai:dspace.mit.edu:1721.1/163028