Home / Current Issue / Paper 1711930
Design and Implementation of a Multi-Modal, Agentic Code Companion System (ICC) using Fine-Tuned Code LLMs and Structured Visualization
Subject area: Science,Engineering and Technology · Area of research: Education
Abstract
This paper proposes the architectural design and implementation strategy for the Intelligent Code Companion (ICC) system, a novel platform engineered to process multi- modal code queries (text and image) and generate comprehensive, structured outputs. The core innovation lies in a hybrid LLM Agent architecture that coordinates specialized tools, specifically Optical Character Recognition (OCR) for handling photographed code inputs and a Structured Output Generator for creating dynamic, animated flowcharts. The system deviates from using underpowered base models, advocating instead for the instruction fine- tuning of a high-performing yet resource- efficient model, such as the StarCoder2-3B model, utilizing Quantized Low-rank Adaptation (QLoRA).
Keywords
Large Language Models, Code Generation, Multi-Modality, Optical Character Recognition (OCR), Agentic Systems, Structured Output, QLoRA Fine-Tuning, Full- Stack Architecture, FastAPI, Perplexity UI.
References
[1] Perplexity AI. (2025). Perplexity AI: The Answer Engine Philosophy.
[2] BigCode Project. (2024). StarCoder2: The Next Generation of Code LLMs. arXiv preprint.
[3] Black, S. et al. (2022). The Stack: A Large- Scale Dataset of Permissively Licensed Source Code. arXiv preprint.
[4] LangGraph Team. (2024). Agent Architecture: Planning and Memory in LLM Systems.
[5] Chen, B. (2023). Retrieval-Augmented Generation (RAG) for Contextual Code Synthesis. IEEE Software.
[6] OWASP Foundation. (2023). A Guide to Parameterized Queries and SQL Injection Prevention.
[7] Dettmers, T. et al. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. arXiv preprint.
[8] NVIDIA & BigCode. (2024).
[9] StarCoder2 Performance Benchmarks: Small Model Viability. Technical Brief.
[10] Baidu Research. (2023). PaddleOCR: Deep Learning for Layout-Aware Text Recognition.
[11] Alqahtani, S. et al. (2023). The Impact of AI on Personalized Learning. Journal of Educational Technology.
How to cite this paper
@article{1711930,
author = {Kunal Kumar Nath, Samridh Chauhan, Tanmay Jadhav, Pranit Virkar},
title = {Design and Implementation of a Multi-Modal, Agentic Code Companion System (ICC) using Fine-Tuned Code LLMs and Structured Visualization},
journal = {Iconic Research And Engineering Journals},
year = {2025},
volume = {9},
number = {5},
pages = {659-669},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1711930.pdf},
abstract = {This paper proposes the architectural design and implementation strategy for the Intelligent Code Companion (ICC) system, a novel platform engineered to process multi- modal code queries (text and image) and generate comprehensive, structured outputs. The core innovation lies in a hybrid LLM Agent architecture that coordinates specialized tools, specifically Optical Character Recognition (OCR) for handling photographed code inputs and a Structured Output Generator for creating dynamic, animated flowcharts. The system deviates from using underpowered base models, advocating instead for the instruction fine- tuning of a high-performing yet resource- efficient model, such as the StarCoder2-3B model, utilizing Quantized Low-rank Adaptation (QLoRA).},
keywords = {Large Language Models, Code Generation, Multi-Modality, Optical Character Recognition (OCR), Agentic Systems, Structured Output, QLoRA Fine-Tuning, Full- Stack Architecture, FastAPI, Perplexity UI.},
month = {November},
doi = {https://doi.org/10.64388/IREV9I5-1711930}
}