01 / Adversarial
Test the combined request
Evaluate text, image, image-editing, and multi-reference workflows for context loss, bypass behavior, composite risk, and guardrail failure.
Adversarial evaluation · Multimodal safety · Quality operations
Jacqueline
Jiang
AI Safety Analyst and red-team operations lead focused on adversarial evaluation, multimodal model behavior, and evaluation quality. I design tests, trace failure mechanisms, and turn recurring errors into review systems that hold.
Evidence / Operating Scale
Operational Record
The public work below shows how I think. These are selected signals from how I have operated: high-volume integrity review, adversarial testing, evaluator QA, root-cause analysis, and quality-system design.
250+
QA audits / week
30%
Team error reduction
800+
Reviews / week
40+
Analysts trained
02 / Quality
Find the mechanism, not just the miss
Separate outcome, evaluator judgment, and failure mechanism; then use recurring errors to drive calibration, secondary review, and remediation.
03 / Systems
Make judgment reviewable
Convert ambiguous safety work into taxonomies, rubrics, escalation paths, and evidence standards that another evaluator can reproduce.
Representative public metrics and methods only. Proprietary datasets, prompts, policies, and internal evaluation materials are excluded.
Applied Analysis
Selected Case File
Synthetic evaluation documenting a model failure across a full behavioral trajectory—not a single isolated turn.
Reusable Methods
Evaluation Frameworks
Methods for classifying intent, diagnosing model-response failures, and locating where multimodal understanding broke.
Analyst Notes
Field Notes
Short analysis on safety judgment and the failures hiding between refusal and compliance.
Observed Behavior
Failure Nodes
Qualitative observations used to isolate recurring behavior patterns worth testing more rigorously.
Analyst Profile
About
I evaluate how generative AI systems interpret intent, context, and risk—especially when meaning emerges across multiple turns or inputs rather than from one trigger phrase.
My work sits at the seam between adversarial testing and operations: design the evaluation, inspect the judgment, isolate the failure pattern, and build the quality layer that keeps the same error from quietly becoming normal.
- Adversarial test design
- Multimodal safety evaluation
- Prompt-injection analysis
- Taxonomy and rubric design
- Evaluator calibration
- Root-cause remediation