Exploring AI-Assisted Development: A Journey with My Side Project
Last weekend I spend building a hobby project. I decided to build it completely using AI supported tools. You can also read it on my blog.
I decided to dive back into machine learning through a hobby project, leveraging AI tools for development and deployment. Here's what I learned along the way.
Development Environment and Tools
AI Assistant Setup
I primarily worked with aider-chat and Anthropic's Claude model. Due to API rate limits, I switched to openrouter for Claude access. While Ollama was an option, its limited context window proved challenging - even with the --map-tokens 1024 parameter adjustment to make the repo map work in aider.
IDE Evolution Initially, I stuck with neovim using avante.nvim, being a long-time vim user. However, I eventually migrated to VSCode for its more polished AI integration. For the moment in VSCode am using cline but be careful using extensions. Cline's recent update introduced some unexpected issues, so it's best to proceed with caution and monitor its performance closely. While codecompanion.nvim looked promising, it wasn't quite ready for prime time. Maybe I need to spend more time configuring it properly.
Technical Hurdles
A particularly frustrating issue emerged from running Aider in PowerShell, which introduced Windows line endings. This caused problems with my Sagemaker oncreate scripts (#!/usr/bin/bash^M). Interestingly, AI couldn't help here - good old-fashioned log debugging saved the day.
Infrastructure Setup
AWS Architecture The project required Sagemaker, IAM, and S3 setup. Aider's /architect mode proved invaluable for initial design planning. It even helped me whip out some diagrams using puml. I leveraged Aider's knowledge graph to research best practices for deploying machine learning models on Sagemaker, ensuring I followed recommended architectures and security protocols. But to be honest I will readup on this using AWS documentation. Ah mentor of mine used to say "trust but verify". When using AI for complex tasks like infrastructure setup, it's crucial to double-check the generated output and ensure it aligns with your specific requirements and best practices.
IAM Challenges Managing IAM roles and permissions was tricky. The AI struggled with IAM's complexity just as a human would, especially when setting up separate data access for different users. While it generated clean IAM and S3 loops, we had to fix some redundant list definitions. Ah very junior developer mistake! It highlighted the importance of careful review even with AI assistance. There were times I felt like I was teaching the AI about IAM as much as learning from it. Downside it will forget it in the next session.
Application Development
Streamlit Implementation I built a straightforward Streamlit app for data upload and inference, outputting prediction CSVs. The AI assistance here felt like working with a junior developer - requiring constant guidance and context-setting. While it helped with features like authentication and S3 bucket integration, the hand-holding needed made me question its efficiency compared to working with human developers.
Debugging and Testing AI excelled at generating detailed debug statements - a preference I have over visual debuggers. While I'm not a TDD enthusiast, AI efficiently generated initial test suites that needed minimal adjustments. After I wrote the code. It was almost like having an extra pair of eyes reviewing my work, catching potential issues before they became major problems.
Reflections on AI in Development
The experience highlighted that while AI can accelerate certain aspects of development, it's not a magic solution. It's most effective for: - Generating boilerplate code - Writing debug statements - Creating initial test suites - Quick function implementations
However, it still requires solid programming fundamentals and constant oversight. While it might reduce team sizes through efficiency gains, complete replacement of developers seems unlikely in the near term.