@mirzayasirabdullahbaig3I built LampLighter an AI-powered software factory for the WeAreDevelopers × BAND Dark Factory Hackathon, hosted on LabLab.ai.
What if AI agents could do more than generate code? What if they could plan, build, test, and independently verify their own work?
That’s the idea behind LampLighter.
The workflow has three key stages:
- Coordinator: Plans and organizes the task.
- Builder: Generates code and tests.
- Verifier: Re-runs the tests on a clean copy and accepts or rejects the result.
Built with Python, Streamlit, and Groq, LampLighter explores how specialized AI agents can work together to make software development more systematic and reliable.
This hackathon was an opportunity to explore agentic AI, multi-agent collaboration, automated testing, and evidence-based verification.
My biggest takeaway: generating code is only part of the challenge. Building a workflow that checks whether the code actually works is just as important.
Project submission:
lnkd.in/ddfM5XJ3
GitHub repository:
lnkd.in/dJRgHDVN
Thank you to lablab.ai, WeAreDevelopers , and BAND for creating an opportunity for builders to explore the future of AI-powered software engineering.
Thanks to Hamna Munir , Zartaj Nadeem and Nouman Arshad
I’m excited to keep building and experimenting with autonomous AI systems.
#AgenticAI #ArtificialIntelligence #AIAgents #SoftwareEngineering #Hackathon #GenerativeAI #Python
Comments2
The Verifier re-running tests on a clean copy is the right instinct. That's the part I still do by hand on my apps: AI writes the code, and a big part of my job is testing and rejecting output. How did you stop the Builder from writing tests that only prove its own code passes?
Running the Verifier on a clean copy catches environment leakage, but the Builder still writes both the code and the tests. Have you tried giving the Verifier its own tests from the spec, or a different model on Groq, to see if rejection rates change?
2 comments
The Verifier re-running tests on a clean copy is the right instinct. That's the part I still do by hand on my apps: AI writes the code, and a big part of my job is testing and rejecting output. How did you stop the Builder from writing tests that only prove its own code passes?
Running the Verifier on a clean copy catches environment leakage, but the Builder still writes both the code and the tests. Have you tried giving the Verifier its own tests from the spec, or a different model on Groq, to see if rejection rates change?