Hi,I'mDonghao 👋
McGill CS · Class of 2027LLM Research @ MilaMontreal, CA
I'm currently a 3rd year Computer Science undergrad at McGill University with experience in fullstack development and AI systems. I'm also a research intern at Mila, working on interpretability for unified vision-language models.
Previously built end-to-end software at Digitech Payments and Group Imi, from a hybrid RAG system for customer support to an AI-powered video production platform.
Interests
Building fast, user-centric tools that make everyday tasks easier
Mechanistic interpretability and representation engineering in large models
Reinforcement learning for post-training and reasoning
Always learning and experimenting with new technologies!
Outside of programming: cooking for my girlfriend, editing video for non-profits, and eating every yummy food in the world 🍜
PyTorch·TypeScript·React·Next.js·FastAPI·Docker·AWS·LangChain·PostgreSQL·Python·SLURM·Redis·PyTorch·TypeScript·React·Next.js·FastAPI·Docker·AWS·LangChain·PostgreSQL·Python·SLURM·Redis·PyTorch·TypeScript·React·Next.js·FastAPI·Docker·AWS·LangChain·PostgreSQL·Python·SLURM·Redis·
Activation Steering·RAG·Multi-Agent Systems·LLM Evaluation·Embeddings·Compiler Design·REST APIs·CI/CD·HPC·MCP·Activation Steering·RAG·Multi-Agent Systems·LLM Evaluation·Embeddings·Compiler Design·REST APIs·CI/CD·HPC·MCP·Activation Steering·RAG·Multi-Agent Systems·LLM Evaluation·Embeddings·Compiler Design·REST APIs·CI/CD·HPC·MCP·
01 / Experience
Mila – Quebec AI Institute
LLM Research Intern
- —Built a multi-gpu inference pipeline with PyTorch for activation steering experiments on unified multimodal LLMs, executing large-scale evaluation sweeps across concepts, prompts, and scaling parameters (180k+ samples).
- —Engineered a 133x performance optimization (from 400+ hours to <3 hours) for autoregressive image generation by architecting a batched, data-parallel pipeline on a SLURM-managed HPC cluster, leveraging KV caching and preallocated attention masks.
- —Implemented low-level activation steering via PyTorch forward hooks, injecting training-free steering vectors into transformer hidden states at inference time for fine-grained control of image generation attributes (e.g., age, emotion, color).
PyTorchDockerHuggingFaceMulti-GPULLM Evaluation
Digitech Payments
Software Developer Intern
- —Shipped a GenAI + hybrid RAG-powered tool over the company's knowledge base to assist a 25+ member customer support team on live calls. Cut average resolution time by 20%. Built with LangChain, pgvector, Electron (React), and Express.
- —Engineered a data ingestion pipeline processing 10k+ internal documents/month using LLM chunking, batched embeddings, and asynchronous processing. Deployed on Google Compute Engine with Terraform, Docker, and GitHub Actions.
Electron.jsExpressLangChainpgvectorDockerGCPPythonTypeScript
Group Imi
Software Developer Intern
- —Led the development of a fullstack video production platform integrating video captioning and multimodal LLM pipelines for highlight segmentation, reducing manual footage review time by 85%. Demoed at Vancouver Web Summit (15k+ attendees).
- —Developed a Claude Sonnet-powered microservice using AWS Bedrock for AI content generation serving 5K+ user requests daily, implementing tool use, prompt templates, streaming responses, fault-tolerant retries, and rate limiting.
- —Implemented 20+ production REST APIs with Express + TypeScript supporting payments, auth, and image/video generation, deployed via CI/CD (Docker + GitHub Actions) to AWS EC2 with nginx and SSL termination.
- —Owned production reliability, diagnosing and resolving issues through logging, monitoring, and systematic debugging.
Next.jsExpressTypeScriptAWS BedrockPostgreSQLAWSPython
02 / Projects
Thread Lens
01LangGraph · MCP · Next.js · FastAPI · Redis · PostgreSQL
- —Developed a multi-agent deep research system that processes and analyzes large-scale Reddit discussion data, decomposing queries into research tasks executed by parallel worker agents and synthesizing findings into structured reports.
C Compiler
02Java · Compiler Design · MIPS Assembly
- —Built a mini C compiler in Java targeting MIPS assembly, implementing a handwritten lexer, recursive-descent parser, semantic analyzer, code generator, and register allocator. Validated against 450+ unit and integration tests across all compilation stages.
03 / Skills
AI/ML
PyTorchScikit-LearnHugging FaceLLM Post-Training (SFT, RL, PEFT)EmbeddingsRAG
Programming Languages
PythonTypeScriptJavaScriptJavaC/C++SQL (Postgres)BashHTML/CSS
Frameworks
Node.jsExpressFastAPIReactNext.jsLangChain/LangGraph
Tools
GitGitHub ActionsDockerTerraformAWSGoogle Cloud PlatformLinux