the work

independent project · python, ollama, faiss

An on-device code reviewer that reads git diffs and returns structured findings from a local 7B model. To get unambiguous ground truth, I mutate clean functions with one known operator and review both versions. The ablation tested whether retrieving past bug-fixes into the prompt helps. It doesn't: recall fell from 40% to 33%, F1 from 0.39 to 0.31. Recall was never the bottleneck anyway, since the model flagged correct code 53% of the time. Published as a negative result.

technical article · medium

Iframe embeds cost 27 to 66 third-party requests per page, at 100KB to 2MB each. The article proposes a Shadow DOM alternative with an explicit embed contract, and pulls apart two things that usually get conflated: the encapsulation boundary and the security sandbox. Shadow DOM replaces the first one and not the second.

30-page case study · Ethical Software Practices

A breakdown of the 2017 Equifax breach and how an unpatched Apache Struts vulnerability (CVE-2017-5638) turned into a full compromise once the organizational failures stacked on top of it: security ownership split across teams, 143 days before anyone detected the intrusion, and access controls that did nothing to contain it. The conclusion was that the missed patch alone would not have been enough. It needed the process failures around it.

independent analysis · responsible AI

A look at bias in judicial datasets and how algorithmic sentencing recommendations widen existing disparities when nobody can inspect why a score came out the way it did. The argument is that a system influencing a sentence has to be interpretable enough for a person to challenge it, not just accurate.