This article is part of a small blog series called Junior-Dev-GPT where I explore ideas to turn LLMs into autonomous junior developers.
Articles: Part 1 | Part 2 | Part 3
Junior Dev GPT Overview
This article is a brief technical overview of Junior Dev GPT. In the video below, I explain the core issue that I am addressing with this proof of concept, along with various use cases that I have discussed with different companies. Additionally, I showcase the interface of Junior Dev GPT when it is running on a local computer.
Current Architecture
The provided illustration depicts the present architecture of Junior Dev GPT. The design choices I have made for this architecture are elaborated in the following video. Moreover, I present a demonstration of the tool’s actual implementation using Langchain in Python.
Future Ideas
This diagram presents my future vision for the architecture of Junior Dev GPT. As mentioned in the video below, there are still several challenging issues that I aim to address in the future. The primary obstacle lies in the noise within the context window, which could be mitigated by effectively organizing the code base into useful chunks without any loss, along with implementing an efficient search mechanism to populate the context window solely with relevant code. Additionally, I plan to experiment with the Coder by incorporating concepts such as a memory stream or a skill/code library, as demonstrated in other research papers. Lastly, I can enhance the depth of the natural language feedback by supplementing it with visual assessments conducted by GPT-4V alongside human feedback.
