Senbao Z.

Key Skills

Software

Internal/Proprietary Tooling

Top Subject Matter

3D multimodal few-shot class-incremental learning (point clouds)

LLM model compression and PEFT fine-tuning for text tasks (detoxification, sentiment regression)

Multi-scale pedestrian detection (image-based) with transformer architectures and supervised training

Top Data Types

3D Sensor

Text

Image

Top Task Types

Fine-tuning

Freelancer Overview

Research Assistant, Tianjin University of Technology (3D Vision and Continual Learning Group) — Few-shot class-increment. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include HuggingFaceTransformers, PyTorch, and PEFT (LoRA). Education includes Doctor of Philosophy, Skolkovo Institute of Science and Technology (2025) and Master of Engineering, Jiangsu University of Science and Technology (2025). AI-training focus includes data types such as 3D Sensor, Text, and Image and labeling workflows including Fine-tuning.

Labeling Experience

PhD Research, Skoltech — LLM Compression for Efficient Deployment with KV-cache compression, quantization, and LoRA-based PEFT fine-tuning.

TextFine-tuning

Conducted LLM compression research focused on reducing memory footprint and latency during autoregressive inference using training-time and evaluation protocols. Applied post-training quantization and LoRA-based PEFT to adapt models for multilingual text detoxification tasks. Fine-tuned BERT for sentiment regression and assessed stability and generative quality under compression constraints. • Performed KV cache compression research (e.g., quantization/eviction strategies) for efficient inference. • Applied post-training quantization and LoRA PEFT for multilingual text detoxification using MT0. • Fine-tuned BERT for sentiment regression tasks. • Built evaluation protocols beyond accuracy to preserve reasoning stability and generative quality.

2025 - Present

Collaboration with Tianjin University of Technology — M3D-FSCL: Multimodal 3D Few-Shot Class-Incremental Learning with CLIP fusion, TLS, and LoRA + distillation.

3D SensorFine-tuning

Developed a multimodal 3D few-shot class-incremental learning framework that trains on labeled 2D depth/image features and 3D point cloud representations. Implemented a CLIP-based 2D–3D feature fusion approach with an EPCL point cloud encoder to support incremental task learning. Introduced a task-adaptive learning strategy and used LoRA fine-tuning with knowledge distillation to balance old and new knowledge during training. • Built a CLIP-based multimodal fusion framework for 3D point cloud FSCIL using depth image features and EPCL encoder outputs. • Added task-adaptive learning (TLS) to improve incremental learning behavior across tasks. • Applied LoRA fine-tuning together with knowledge distillation for retention of previously learned classes. • Published a first-author paper in MultimediaSystems (JCR Q1).

2024 - 2025

Research Assistant, Tianjin University of Technology (3D Vision and Continual Learning Group) — Few-shot class-incremental learning on 3D point clouds with multimodal CLIP-based framework and LoRA fine-tuning.

3D SensorFine-tuning

Investigated few-shot class-incremental learning for 3D point clouds and developed training frameworks to improve robustness to continual learning settings. Implemented LoRA fine-tuning for a point cloud encoder and used task-adaptive learning strategies to reduce catastrophic forgetting. The work required preparing and using labeled 2D depth images and 3D point cloud features as model inputs for supervised/continual training. • Built a CLIP-based multimodal framework (M3D-FSCL) combining 2D depth images and 3D point cloud features for FSCIL. • Trained and adapted a point cloud encoder using PEFT/LoRA for task-specific learning. • Designed task-adaptive learning to balance old vs. new knowledge during incremental updates. • Published results in MultimediaSystems (JCR Q1) as first author.

2024 - 2025

Master Thesis & Jiangsu Provincial Project — Multi-scale pedestrian detection using GPP augmentation, ViT-CA, and scale-weighted IoU training.

ImageFine-tuning

Designed and trained an object-detection pipeline for multi-scale pedestrian detection using vision transformer-based feature fusion. Developed grid-based probabilistic pasting (GPP) data augmentation and proposed a scale-weighted IoU loss to better supervise small-scale targets. Created and used training objectives and augmentation strategies that effectively depend on labeled pedestrian bounding regions/targets in datasets. • Implemented GPP data augmentation and a Vision Transformer with Channel Attention (ViT-CA) for multi-scale feature fusion. • Proposed a scale-weighted IoU loss to improve small pedestrian detection performance. • Trained and evaluated on Caltech and CityPersons datasets to achieve state-of-the-art results. • Published in Expert Systems with Applications (JCR Q1) as part of a Jiangsu provincial innovation project.

2023 - 2025

Education

Work History