Johns Hopkins SAIS. Building reliability tooling for AI systems. Everything below ships with deterministic verification built in, not bolted on.
fieldatlas: automated literature review for AI safety where every claim is verified against source text and every citation is linted against a real corpus. Ground-truth tested on 15 papers blind to the system: 15/15 faithful, 0 fabrications, 0 material errors. Generalized to a second field (CRISPR governance) on the first run.
nla-eval: interpretability evaluation of released NLA checkpoints on H100. Published negative results honestly. Zenodo DOI.
| Project | What it does | Evidence |
|---|---|---|
| aslan-core | Bitemporal financial-data platform (Turkish market). Immutable history, content-addressed provenance, LLM extraction audit trail | 29K lines, 9 releases, ~1000 tests |
| gpuwatch | Agent-safety study: purchase agent with hard-coded spend caps + HMAC confirmation gate that config cannot override | 41 security tests passing |
| dc-frontier-events | DC AI and frontier-tech event aggregator. Multi-source dedup and relevance ranking | Live at events.emersus.ai |
| emersusai | RAG health assistant over 1.1M biomedical papers. Refuses rather than hallucinates | 1,345 commits in 6 weeks solo (archived) |
| h1b-labor-map | Federal-data ETL: which employers sponsor H-1B, where, in which roles. Every threshold cited to the Federal Register | Ran on real multi-GB federal releases |
Turkish (native), English (C2). ORCID.