Chris Deschenes

I am a research engineer and founder focused on empirical safety and evaluation for agentic AI systems. My background spans three decades in applied machine learning and software engineering, including computer-vision systems deployed in HIPAA-, SOC 2-, and FDA-constrained clinical workflows, as well as cloud-based multi-agent systems that use tools, maintain state, and operate behind human approval, permission, kill-switch, and audit controls. My recent work emphasizes outcome-grounded evaluation: building deterministic and model-based graders, analyzing agent traces, testing calibration under distribution shift, and using operational evidence to distinguish genuine capability from benchmark artifacts or persuasive but unsupported outputs. In one multi-agent decision system, a forensic evaluation rejected the hypothesized model edge and exposed execution failures larger than the measured gains. In a separate probabilistic forecasting project, a recalibration method that improved synthetic results degraded under a real regime shift. I am especially interested in evals, oversight, and AI control for long-horizon agents, including deceptive completion claims, unsafe tool use or delegation, approval-boundary violations, reward hacking, evaluator gaming, and reliable escalation or shutdown.

Personal Website

https://relightlabs.ai/

LinkedIn

LinkedIn

Projects

No grants recorded.