selected work

Four projects, and the thing they have in common is that each one should be able to show why it produced what it produced.

Code Review Agents is the largest of them and the one with a paper behind it. Its reviewers cannot report a finding unless they can trace it to tool output or a named line. Self-RAG carries the same rule into retrieval, so an answer is graded against the documents meant to support it and one that is not grounded there is withheld. ReAct Agent is the smallest, and I wrote it to make the loop legible.

GPT-2 Turkish is the odd one here. It is the oldest of the four, and a pretrained base model, not a system. I published it in 2022 for people who needed a Turkish checkpoint to fine-tune and did not have the hardware to train one themselves.

All four are open, and the ones with code keep their tests and their evaluation harnesses in the same repository as the code.