Dong
Luanjie

Worked in TikTok's Global E-commerce Knowledge Graph & AIGC team, fine-tuning multimodal LLMs for product classification.

• Re-architected a single-stage classifier into a two-stage recall-and-ranking framework with multimodal LLMs fine-tuned via SFT and RL (DPO/GRPO) — +10% accuracy over production
• Designed a data strategy leveraging raw e-commerce signals (image, title, attributes) to build high-quality training samples, adding a further +5% accuracy
• Built a multi-agent training pipeline with context and harness engineering, cutting training iteration cycles by 50%

Tech Stack

Python

Pytorch

Worked in KPMG's Forensic Team optimizing and deploying machine learning pipelines for forensic data analysis.

• Redesigned feature engineering pipeline using Pandas vectorization — reduced processing from 5 minutes to 5 seconds (98% improvement)
• Developed and trained a Generative Adversarial Network (GAN) to generate synthetic datasets, yielding 10% accuracy increase
• Containerized ML models using Docker and deployed on Microsoft Azure with Terraform IaC

Tech Stack

Javascript

HTML

CSS

React

Python

MySQL

Docker

Azure

Pandas

Scikit-learn

Working with Team 4/5 of SMART M3S, focusing on AI-driven solutions for urban mobility, economic growth, and human-machine collaboration.

• Building NLP extraction pipelines to analyze global firms' annual reports and extract key research metrics
• Conducted spatiotemporal analytics on GPS mobility data using KMeans clustering to identify commuter profiles
• Detected patterns in Points of Interest using Apriori and Edit-Distance metrics
• Supporting academic research on mobility behavior and sustainable city planning

Tech Stack

Python

Pytorch

TensorFlow

Pandas

Scikit-learn

Developed an end-to-end MVP counseling platform named AImpathy, integrating multi-modal deep learning pipelines to provide actionable insights from therapy sessions.

• Designed a scalable video ingestion pipeline using AWS S3 for storage and AWS SQS for event-driven processing
• Engineered a Transformer-based multi-modal system combining Whisper ASR for speech transcription and custom models for emotional analysis
• Containerized ML models using Docker and deployed on AWS ECS for scalable performance
• Leveraged OpenCV for preprocessing visual cues from recorded sessions

Tech Stack

Javascript

HTML

CSS

WordPress

PHP

Python

Docker

AWS

Pytorch

TensorFlow

Pandas

Scikit-learn