This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
I created Sloth-Compile V2 to make fixing Python errors less frustrating. Instead of repeatedly reading tracebacks, copying code into a chatbot, and comparing changes manually, I can drop a Python file into the sloth_dropzone folder. I can also include a screenshot of an error message. The application routes text-only requests to Gemma 4 and
screenshot-assisted requests to Phi-3 Vision, according to my current configuration. It saves proposed fixes separately in sloth_processed,
leaving the original file untouched. SQLite stores processing history and code differences, while a Streamlit dashboard lets me review files and inspect unified diffs. Fixes still need human review and testing.
Demo
https://drive.google.com/file/d/1CmJeZMCe3TO1WyIjufuntbyYPTAx0vYM/view?usp=sharing
Code
Sloth-Compile V2: Multimodal Drop-Box Edition
An intelligent local Python debugger and code-tweaking agent powered by Ollama.
Features
- Background drop-box watcher for automated code repair
- Multimodal error traceback detection using local vision models
- In-memory code tweaker and Q&A workspace via Streamlit dashboard
- Sandboxed execution verification
How I Built It
I built Sloth-Compile with Python, Ollama, Watchdog, SQLite, Streamlit, and difflib. Gemma 4 (gemma4: e2b) handles text-based debugging, while Phi-3 Vision (phi3: vision) handles requests that include screenshots. Both models are called through Ollama’s local /api/chat endpoint. Watchdog monitors the drop-box folder and waits briefly for a matching screenshot. Separate modules manage configuration, model requests, folder monitoring, database history, and code comparisons. SQLite stores original code, proposed repairs, diffs, timestamps, status, and model details. Streamlit provides a local dashboard to review results. Generated fixes are proposals, not proof that code works.
Why Does Open Innovation Matter?
Open innovation let me experiment with an open-weight model running locally instead of relying entirely on a hosted AI API. Using Gemma 4
through Ollama gave me a way to connect model responses to a folder watcher, database, file-processing workflow, and dashboard. This made
the project more than a chat interface. The original code remains separate from the proposed repair, and users can inspect the differences before deciding whether to use them. Local inference also means source code does not need to be sent to a third-party AI API as part of this workflow, although other services and configurations can affect privacy.
I can change models, prompts, and components as the project develops.
The Lazy Payoff
Sloth-Compile aims to reduce repetitive steps in Python debugging. Instead of manually copying code and tracebacks into an AI chat, copying the answer back, and comparing files, I can drop a file into a folder and review the suggested fix in the dashboard. Adding an error
screenshot can provide extra context without manually transcribing the message. The history also lets me revisit earlier attempts, see which model was used, and inspect the changes. Because the original source is preserved, I can review the proposal before adopting it. The actual time saved depends on the bug, model speed, and correctness of the response. I have not measured time savings yet; testing known bugs and recording pass rates would help evaluate the benefit.
Prize Categories
Sloth-Compile fits several project areas: open-weight models, local inference, AI agents and developer tools, multimodal AI, and open-source
tooling. It uses Gemma 4 for text-based debugging, Ollama for local model access, and a separate vision model for screenshot-assisted
debugging. Python, Watchdog, SQLite, Streamlit, and difflib connect the model to a practical workflow. These are descriptive areas, not verified official DevRelay prize categories. Before submitting, I should compare them with the challenge’s official rules and use only categories that actually exist and apply. If a Gemma or Google AI category is available,
I can consider it if the rules permit it. The project’s central idea is turning open-weight AI into a reviewable local debugging tool.
Top comments (0)