Wenhe Wang

Hello! 👋

I'm Wenhe Wang, a Senior Applied Machine Learning Scientist with a Master's in Visual Computing from SFU. I build production AI systems across LLM, RAG, and agentic pipelines, computer vision, and cloud data infrastructure — shipping models for fraud detection, content moderation, image processing, multi-agent assistants, and conversational AI.

My work spans sectors from CAD, BIM, edge devices, fintech, and consulting to short-video platforms, including leading projects at Autodesk, Samsung Research Canada, Cavallo Technologies, TikTok, China Guangfa Bank, and KPMG.

I received my M.Sc. in Financial Engineering from University of Reading in 2019 and my B.Ec. in Financial Engineering from Nankai University in 2017.

Projects

Selected personal, research, and course projects across NLP, computer vision, cloud systems, and finance.

Natural Language Processing

Preference Optimization

Wenhe Wang, Gouttham Nambirajam, and Raj Shreyas Penukonda
SFU Class Project, CMPT 713: Natural Language Processing

Project Page

Improved instruction-following behavior of Qwen 2.5 0.5B-Instruct by fine-tuning with the ORPO preference-optimization framework as an alternative to RLHF.

Prompt Tuning for Text Generation

Wenhe Wang, Gouttham Nambirajam, and Raj Shreyas Penukonda
SFU Class Project, CMPT 713: Natural Language Processing

Project Page

Adapted GPT-2 to table-to-text generation via lightweight prompt tuning, demonstrating parameter-efficient adaptation without full fine-tuning.

BERT Finetuning for Robust Phrasal Chunking

Wenhe Wang, Gouttham Nambirajam, and Raj Shreyas Penukonda
SFU Class Project, CMPT 713: Natural Language Processing

Project Page

Built a robust phrasal chunker by fine-tuning a BERT-based Transformer, comparing transformer transfer learning against classical sequence-labeling baselines.

Word Vectors and the Analogy Task

Wenhe Wang, Gouttham Nambirajam, and Raj Shreyas Penukonda
SFU Class Project, CMPT 713: Natural Language Processing

Project Page

Solved word-analogy tasks with pre-trained GloVe embeddings, then improved their semantic structure via retrofitting against external lexical resources.

Segmentation into Words

Wenhe Wang, Gouttham Nambirajam, and Raj Shreyas Penukonda
SFU Class Project, CMPT 713: Natural Language Processing

Project Page

Recovered word boundaries from unsegmented English text using an N-gram language model, returning the most plausible segmentation under a probabilistic scoring function.

Computer Vision & Graphics

Unity Cyberpunk Hologram and Teleport effects

Wenhe Wang and Gouttham Nambirajam
SFU Class Final Project, CMPT 766: Computer Animation

Final Effect, Presentation Slides

Reimplemented two Cyberpunk-inspired 3D visual effects — a hologram and a teleport — from scratch in Unity using custom shader scripts and the Unity Render Pipeline.

3D Reconstruction using Neural Fields (NeRF)

Wenhe Wang
SFU Class Project, CMPT 743: Practices in Visual Computing 2

Project Page

Reconstructed 3D objects from occupancy-field samples by training an MLP over a learned 3D feature grid — a neural-field alternative to mesh-based reconstruction.

Latent Manipulation with StyleGAN2

Wenhe Wang
SFU Class Project, CMPT 743: Practices in Visual Computing 2

Project Page

Controlled generated image attributes via StyleGAN2's latent space — combining interpolation, style mixing, and directional manipulation to seamlessly blend and alter visual features.

Image to image translation with CycleGAN

Wenhe Wang
SFU Class Project, CMPT 743: Practices in Visual Computing 2

Project Page

Translated between unpaired image domains (Apple ↔ Windows emojis) by re-implementing CycleGAN from scratch, preserving structural content while transferring style without paired training data.

Pose Regression with PoseNet

Wenhe Wang
SFU Class Project, CMPT 743: Practices in Visual Computing 2

Project Page

Regressed 6-DoF camera pose from single images by re-implementing PoseNet with a modified InceptionV1 backbone pre-trained on the Places dataset.

Object Detection with SSD

Wenhe Wang
SFU Class Project, CMPT 743: Practices in Visual Computing 2

Project Page

Built a simplified Single-Shot MultiBox Detector with multi-scale feature maps and anchor priors, contrasting the SSD approach against YOLO-style single-grid detection.