Debugging LLMs Like Code? I Built an AI Autopsy Engine

Description

Most AI observability tools stop at prompts and responses. I wanted to go deeper. In this video, I break down Synapse-Graph — an AI Autopsy Engine that turns LLM internals into observable and governable infrastructure. Instead of treating models like black boxes, this system: - Traces attention heads using PyTorch forward hooks - Maps internal flow as a lineage graph - Runs causal discovery via ablation sweeps - Quarantines defective circuits at runtime (no retraining) Core idea: Treat LLM internals like a data platform. model → database layers → tables attention heads → columns activation flow → lineage ⚙️ Tech Stack: Python, FastAPI, PyTorch, Transformers Next.js, React, Tailwind OpenMetadata, Ollama 🚀 Repo: https://github.com/FiscalMindset/Synapse-Graph 📌 What this answers: “Did this specific internal component cause this specific output?” That’s where real AI engineering starts. --- If you're building in AI, LLMs, or ML infra — this is worth exploring. #AI #LLM #MachineLearning #OpenMetadata #FastAPI #DeepLearning