Bingxi J.

Key Skills

Software

CVAT

Label Studio

Top Subject Matter

Computer vision for vehicle damage assessment and defect detection

LLM representation evaluation and prompt robustness testing

Top Data Types

Image

Text

Audio

Top Task Types

Fine-tuning

Data Collection

Transcription

Evaluation/Rating

Object Detection

Point/Key Point

Classification

Segmentation

Bounding Box

Text Generation

Freelancer Overview

Computer Vision Algorithm Engineer Intern at Ping An Technology (June 2025 – September 2025) focused on VLM training wor. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include LLaMA-Factory and MLflow. Education includes Bachelor's Degree, University of California, Irvine (2028). AI-training focus includes data types such as Image and Text and labeling workflows including Fine-tuning, Evaluation, and Rating.

Labeling Experience

Research Assistant at Relational Cognition Lab, UCI (January 2026 – Present) building LLM evaluation harnesses and experiment orchestration.

TextEvaluation/Rating

Built an evaluation harness to probe and compare model representations across Gemma, Qwen, and Llama using standardized prompt sampling and aggregation. Implemented a noise-estimation baseline by injecting random text into prompts to quantify variance for defensible comparisons. Orchestrated repeatable experiments with an evaluation framework and tracking to ensure reproducibility, then deployed an inference service to serve embeddings for downstream analysis. • Standardized prompt sampling and aggregation for multi-model evaluation • Noise-injection baseline for variance/robustness measurement • Experiment orchestration with reproducible tracking and metrics • FastAPI + Redis embedding inference service to reduce downstream analysis latency

2026 - Present

Computer Vision Algorithm Engineer Intern at Ping An Technology (June 2025 – September 2025) focused on VLM training workflow and production inference pipelines.

ImageFine-tuning

Designed multi-stage vehicle-damage inference workflows using multi-view inputs and prompt-based reasoning to produce assessment reports. Productionized a VLM training workflow by packaging Qwen2.5-VL with LLaMA-Factory, including performance-driven evaluation. Built image-quality enhancement and defect detection components using CLIP semantic signals, diffusion features, and PatchCore framing to improve stability and decision quality. • Vehicle damage assessment pipeline with multi-view ingestion • VLM training workflow packaging and optimization (Qwen2.5-VL, LLaMA-Factory) • CLIP-guided image quality enhancement for stability • Defect detection workflow combining diffusion + CLIP + PatchCore framing extraction and decision service

2025 - 2025

Education

Work History