Skip to content
AegisNet
Back to community

Amara Okafor

ResearcherOpen to collaborate

Interpretability researcher — sparse autoencoders & feature steering

London, UK

About

Working on making model internals legible enough to audit. Previously physics; now decomposing residual streams and publishing replication-friendly tooling.

Projects & outputs

  • FeatureAtlas

    2026

    Open catalogue of ~40k interpretable SAE features with search and provenance.

  • Steering under distribution shift

    2025

    Measured how activation steering degrades outside the calibration set.

Trusted recommendations

Written by people whose organization affiliation has been verified on AegisNet. Candidates choose which recommendations to attach to an application.

  • Amara ships careful, reproducible work and flags her own negative results first.

    Dr. Iris Vandermeer · Meridian Alignment Lab Verified