in progress
AI
2025

AGI/AI Safety Research

Bachelor Thesis: Designing a Self-Improving Research Agent with Formal Verification towards AGI.

Python
PyTorch
LangGraph
Z3
Coq/Lean
HuggingFace Transformers

🎯 The Problem

Current LLM-based autonomous agents lack formal guarantees for safety and correctness, making them unreliable for critical applications. As AI systems become more autonomous and capable, there is an urgent need for provable safety mechanisms and formal verification to ensure reliable operation in AGI development.

💡 The Solution

Developing an LLM-based autonomous research agent with iterative self-improvement capabilities, integrated with formal verification methods (Z3, Coq/Lean). Implemented using LangGraph prototype for recursive research, document/code generation, and feedback-driven optimization. The system combines the flexibility of LLMs with mathematical guarantees of formal methods.

🚀 The Outcome

This bachelor thesis research contributes to practical AGI development with provable safety properties. The work bridges formal verification and modern AI systems, creating frameworks for self-improving agents that maintain safety guarantees. Results will be published and contribute to the field of AI Safety and reliable autonomous systems.

Project Visuals

Check out the GitHub repository for code samples, demos, and detailed implementation notes.

Source Code

Available on GitHub

Documentation

README & Guides

Explore More Projects