Alex Y.

Key Skills

Software

Other

Top Subject Matter

AI training data engineering for coding agent evaluation

LLM/vision-language model evaluation and red-teaming

AI safety / alignment

Top Data Types

Image

Text

Computer Code Programming

Top Task Types

Computer Programming/Coding

Red Teaming

Evaluation/Rating

Question Answering

Freelancer Overview

I'm a contractor with Handshake AI on two active training and evaluation projects for frontier AI labs. On Project Helix, I convert real open-source bug fixes (Python and Rust) into clean evaluation benchmarks for AI coding agents — separating fixes from tests, writing symptom-focused problem descriptions that don't leak the solution, and validating in reproducible Docker environments; I automated much of the pipeline with the Claude Code CLI, cutting per-task time from hours to under 30 minutes. On Project Escher, I design test cases that expose vision-language model weaknesses in spatial reasoning, color, counting, and occlusion, and was promoted to Reviewer to grade other contributors' submissions against a 5-point rubric. I'm a Rice University CS and Operations Research (Financial Engineering) double major (4.00 GPA) and a Rice AI Alignment Fellow, with working knowledge of Python, TypeScript, Rust, Docker, and Git/CI.

Labeling Experience

AI Visual Reasoning Evaluator & Reviewer (Project Escher) — Handshake AI (May 2026 – Present)

OtherText

Designed and reviewed AI visual-reasoning evaluation tasks to measure frontier vision-language model performance. Created test cases featuring charts, diagrams, maps, and infographics targeting spatial reasoning, color discrimination, counting, and partial occlusion. Reviewed peer submissions using a 5-point rubric and flagged low-quality answers early to reduce rework. • Targeted weaknesses such as hidden/overlapping shapes, fine color differences, and multi-panel tracking • Promoted to reviewer after sustained quality • Graded submissions on answer correctness, prompt clarity, copyright, and model failure modes • Achieved elevated model-failure rates by concentrating on known weak spots

2026 - Present

AI Training Data Engineer (Project Helix) — Handshake AI (May 2026 – Present)

Converted real-world open-source bug fixes into reference materials for evaluation of AI coding agents tested on the changes. Prepared bug descriptions centered on observable symptoms with concrete input/output examples rather than revealing implementation hints. Built and automated Docker-based test environments to ensure consistent evaluation across multiple repositories and failure modes. • Separated fixes from test cases into clean, standardized commits • Authored symptom-focused descriptions to force genuine problem solving • Debugged environment issues including OOM errors, broken runners, and misconfigured builds • Automated setup, patch application, and commit structuring via Claude Code CLI

2026 - Present

Education

Work History