Exhibit C
Three years evaluating frontier AI
Since 2023, professional evaluation of the large language models behind today’s AI products — the working knowledge of how these systems behave, and fail, that grounds every AI integration I build.
- Period
- 2023 — present
- Focus
- Code & reasoning evaluation
- Method
- RLHF · adversarial prompting · strict rubrics
LLM evaluation · RLHF · prompt engineering
The work
Since 2023 I have worked as an AI trainer and RLHF specialist for vendors serving frontier AI labs — evaluating and ranking model outputs on code and reasoning tasks, designing adversarial prompts that probe where models break, and reviewing peer submissions under strict quality rubrics. This is the human feedback loop that today’s leading language models are trained on.
Why it matters for your project
Most “AI integration” fails in the gap between a demo and production: the model that looked brilliant in a test hallucinates a price, mangles an edge case, or degrades quietly after a prompt change. Three years of professional evaluation means I design around those failures from the start:
- Eval-first integrations — every AI feature ships with tests that measure whether the model is actually doing the job, not just whether the code runs.
- Failure-mode literacy — I know the difference between what a model can reliably do and what it can only demo.
- Honest scoping — some of your process should be automated with plain software, not AI. I’ll tell you which part is which.
The outcome
AI features built on this foundation do the job they claim — or they don’t ship. That judgment is the difference between AI as a feature and AI as a liability.