Research

What I build, and what I measure.

Speech and language models

At Bhala AI I pretrain speech and language models from scratch for communities AI has excluded — corpus construction, tokenizer, pretraining, fine-tuning, evaluation and deployment.

Model evaluation

I audit open frontier models for what their representations encode about race and gender, reading hidden states rather than outputs. One protocol throughout: linear and MLP probes across sixteen to seventeen layers, twenty-eight protected dimensions, 10,064 sentence pairs, five-fold cross-validation, with random-label and random-feature controls and a norm-matched random-direction floor.

Method

I run the controls that can invalidate the result. A lexical-swap control showed that a separability result I had already published was reading word identity rather than stereotype, and I retracted it. Results are scored by public classifiers trained by other teams, so verifying a claim requires downloading them rather than trusting me. When a pre-committed experiment failed, I published the failure; when a widely used benchmark proved defective across six models, I reported that too.

Every number ships with its data file.