Ryle Goehausen
Engineer working in Rust, Scala, Python and Spark: data systems, performance engineering, machine learning and distributed systems.
Director of Engineering at Constellation Network, where I lead our open-source project and applied-AI work: DoD / U.S. Air Force supply-chain delivery, LLM security and LLM inference optimization. Before that, a senior engineer at Bloomberg working on Parquet-based portfolio data lakes.
Publications & writing
-
Gate AI: LLM Security Benchmark Evaluation Methodology and Results
Evaluation harness for prompt-injection and jailbreak detectors: 16 public benchmarks, leakage-controlled cross-validation and one global operating point at ≤1% false-positive rate.
-
Frontier models fall to attacks you can script
Fixed, replayable injection payloads compromise leading models.
-
Your system prompt is not a security control
A defended email agent, one forged message: instruction-based defenses fail where architectural controls hold.
Recent writing
- Redacting secrets from LLM requests without breaking the prompt cache
- What I learned running a fleet of AI coding agents
- Single-user speed vs total throughput: reading inference-chip claims
- Hardware transcoding Jellyfin on an Intel NUC
- Looking at one token of LLM inference
- What CI costs on Hetzner versus GitHub Actions