Aashu User

I’m an AI Data Specialist who builds dependable evaluation pipelines for LLM, LAM, and CUA systems—turning raw model behavior into actionable findings. I run benchmarking audits against acceptance criteria and evaluation rubrics to surface misalignments, regulatory risks, and clear error ratings that teams can use immediately.

In my recent work, I designed and refined prompt templates for agentic AI workflows, iterating until instruction following, model grounding, and autonomous execution were consistent across test cases. I also validate video trajectory datasets for CUA training by cross-checking spatial and temporal annotations against labeling guidelines to keep training inputs clean.

My projects reflect the same focus on realism and rigor. For CUA-Bench, I built a benchmark dataset of 100+ prompts with conditional branching and multi-constraint scenarios to evaluate decision-making, fallback handling, and instruction following under conditions real users actually run. For my LLM and LAM Failure Analysis and Audit Framework, I used failure taxonomy mapping and behavioral pattern classification to produce annotated datasets and error-rated audit reports.

I take quality seriously: I’ve maintained data accuracy above 99% through shared audit checklists, labeling standards, and quality reviews for teams of annotation specialists. I’m motivated by improving model retraining cycles with reproducible benchmarking frameworks, root cause analysis, and measurable robustness.