I am a Research Engineer bridging the gap between theoretical AI safety and scalable infrastructure. My work focuses on Knowledge Provenance—tracing the specific training data responsible for model behaviors during inference.
I build "Glass Box" tooling to make Large Language Models auditable by design.
- Knowledge Provenance Maps: A framework for tracking gradient influence at the document level during the training loop.
- Mechanistic Interpretability: Developing 3D visualization tools to map "knowledge anatomy" across transformer layers.
- Model Health Metrics: Author of the Sparsity, Concentration, and Utilization metrics for diagnosing fine-tuning efficacy.
- Knowledge Provenance Maps for Enhanced Neural Network Interpretability (2025)
- Granular Knowledge Provenance: Bi-Directional Document-to-Node Mapping (2025)
- Temporal Analysis of Knowledge Structure (2025)
- knowledge-provenance-suite: A PyTorch-based library for tracking training data influence and visualizing 3D knowledge maps. (Formerly "Transparent AI Suite").
- granular-gradient-tracker: Memory-efficient implementation of per-sample gradient tracking for LLMs.
- Collaboration: I am open to research collaborations on interpretability and model steering.
- Email: jonathanbaraldi@gmail.com