About Kris Leigh Miller
I want to know what the evidence actually supports.
I work at the boundary between a system producing an output and producing work that is correct, complete, usable, and safe to rely upon.
What connects the work
My projects look different on the surface: AI benchmarks, agent-assisted software development, semantic retrieval, workflow automation, technical research, and writing. The recurring question is the same: what would we need to observe before relying on this result?
That means defining the real requirement, identifying the acceptance boundary, preserving evidence, investigating failures, and stating what remains uncertain. I am less interested in whether a system can produce something impressive than in whether the result survives inspection and fits the work it was supposed to support.
Background
My background spans psychology, information technology, technical support, education, AI data evaluation, benchmark engineering, agent-assisted product development, technical research, and public writing.
I hold bachelor’s degrees in Psychology and Information Technology. Earlier work in technical support and legal administration taught me to investigate unfamiliar problems, manage detailed records, and follow an issue through resolution. Fourteen years of individualized education strengthened the same skills from another direction: research, adaptation, assessment, documentation, and designing around the person rather than an imaginary average user.
My more recent work focuses on AI quality and reliability. I design evaluations and acceptance criteria, test model and agent behavior, supervise agent-assisted implementation, investigate integration failures, assemble continuity and knowledge systems, and translate technical findings into usable decisions.
How I work
- Inspect before inventing. I examine native capabilities, existing code, and public implementations before recommending custom work.
- Define the acceptance boundary. A passing test, completed build, or polished report is evidence only for what it actually checked.
- Preserve the failures. Contradictions, blocked tests, evaluator mistakes, and unresolved risks belong in the record.
- Keep human authority explicit. Automation may perform substantial work without silently acquiring the right to make consequential decisions.
- Write the reasoning down. Requirements, decisions, corrections, sources, and limitations should remain available for review.
Current focus
I currently work across applied AI quality and reliability, agent evaluation, systems QA, technical investigation, human-supervised workflow design, knowledge systems, and evidence-led technical writing.
I am open to remote full-time and contract roles, as well as tightly bounded evaluation engagements using sanitized workflows and written communication.
Communication
I work remotely from Rockport, Texas. I prefer written communication because it keeps requirements, questions, decisions, and changes explicit and reviewable.