Back to portfolio
Case study / Project archive
NeuroScope
An interpretability interface for exploring model activations and experimental agent-analysis workflows.
Project archive
Historical project; AI-assisted development. Individual and team contribution has not been independently verified.
The idea
The problem
Model outputs alone hide what changed inside a run. Researchers need to inspect activations and test interventions without mistaking correlation for causation.
The implementation
What I built
The inspected source joins run inspection with sparse-feature analysis and intervention tooling. It explicitly labels correlation-based co-activation separately from causal patching.
- Python
- PyTorch
- React
Features
- Inspect model runs, per-step activations, sparse features, and saved experiment findings.
- Explore activation patching, patch matrices, and feature steering through separate API paths.
- Plot sparse-feature timelines and exploratory co-activation relationships.
Architecture
- React experiment interface with FastAPI routers for runs, experiments, steering, and probe training.
- PyTorch model hooks and SAE decomposition feed saved step artifacts; patching and steering modules perform distinct interventions.
Engineering decisions
- Call the Pearson-correlation view a co-activation graph; correlation alone does not establish a circuit's causal role.
Validation and measured results
- Local review covered run and steering routes plus SAE, patching, and steering modules. No model execution or causal finding was independently reproduced.
Limits and unfinished work
- Feature descriptions, drift estimates, and steering outcomes require model-specific experiments; an interface or graph is not proof of interpretability accuracy.
- Older source can contain superseded descriptions. Numerical impact, adoption, awards, and deployment claims are not carried into this entry.
Evidence and sources
This project’s evidence remains local. Public source and downloadable receipts are not attached yet.