张辰龙

Key Skills

Software

Labelbox

LabelImg

Data Annotation Tech

Supervisely

Other

Internal/Proprietary Tooling

Top Subject Matter

query search and text classification

Multimodal image-to-text retrieval

geometry diagram reconstruction

Top Data Types

Text

Image

Top Task Types

Fine-tuning

Text Generation

Emotion Recognition

Bounding Box

Classification

Object Detection

Question Answering

Text Summarization

RLHF

Computer Programming/Coding

Function Calling

Prompt + Response Writing (SFT)

Freelancer Overview

M.S. candidate in Software Engineering with strong hands-on experience in LLMs, multimodal retrieval, text classification, distributed training, and model deployment. Published in SCI journals and experienced in building production-ready NLP and vision-language systems for educational and medical scenarios.

Labeling Experience

NLP Algorithm Engineer (TAL) — Intent recognition and keyword extraction for AI chat

TextFine-tuning

Designed intent recognition and keyword extraction for an AI chat assistant with four intent classes (feature inquiry, casual chat, study materials, and other). Used DeepSeek to distill data and perform imbalance-aware sample synthesis before fine-tuning Qwen2.5-Instruct models on a 300K-sample dataset. Built a real-user evaluation set and measured strong JSON parse reliability and intent classification accuracy.

2024 - 2025

NLP Algorithm Engineer (TAL) — Geometry diagram reconstruction with VLMs

ImageText Generation

Explored VLM-driven geometry diagram reconstruction using structured text generation for geometry images. Created synthetic training data by rendering basic geometric shapes with Matplotlib, then trained Qwen2.5-VL Instruct models on 400K samples. Achieved high reconstruction accuracy on engine-generated validation data while monitoring real-world corpus performance.

2024 - 2025

NLP Algorithm Engineer (TAL) — BLIP2-based image-to-text retrieval

ImageFine-tuning

Built an image-to-text multimodal retrieval system using a BLIP2-based pipeline with a vision encoder and Q-Former to retrieve question text from user-captured images. Adapted the training objective for image search by retaining only image-to-text similarity within the ITC loss and changing retrieval scoring via mean-pooled query tokens. Trained on 2M image-text pairs using DeepSpeed and evaluated retrieval performance with Recall@44 over a large indexed corpus.

2024 - 2025

NLP Algorithm Engineer (TAL) — Photo-based question search subject classification

Don't discloseTextFine-tuning

Performed subject classification for photo-based question search by continuing pretraining Chinese-RoBERTa with masked language modeling on a 20M-sample dataset, then fine-tuned using embedding aggregation and multiple regularization techniques. Applied training-time label smoothing and R-Drop to improve classification robustness. Validated model quality on an evaluation set and reported the resulting accuracy improvements.

2024 - 2025

Project — Intelligent Medical Triage Assistant for an Elderly Care Community

Fine-tuning

Trained a medical triage assistant by preparing domain vocabulary for Chinese medical content and continuing pretraining a Llama-2-7B model with LoRA on medical corpora. Performed LoRA-based instruction tuning on 4M triage dialogues and clinical diagnosis records to align the model with triage and diagnosis dialogue behavior. Deployed the resulting service with vLLM and FastAPI to support streaming responses for the medical demonstration use case.

2024 - 2024

Show more

Education

Work History