Hi, I'm Steven Huang 👋

I'm a junior ML/AI Software Engineer and UNSW Master of Information Technology graduate specializing in Artificial Intelligence. I build practical AI software with Python, FastAPI, Docker, CI/CD, cloud deployment workflows, and RAG/LLM systems.

My strongest direction is AI application engineering: turning ML/RAG ideas into reproducible, testable software with clean APIs, local/containerized workflows, evaluation, and honest documentation of limitations.

• AI / LLM Applications — RAG pipelines, embeddings, FAISS retrieval, citation-aware answers, provider abstraction, and evaluation workflows
• ML-backed Products — data cleaning, feature engineering, model comparison, Optuna tuning, inference APIs, and user-facing applications
• Backend AI Systems — FastAPI services, PostgreSQL-backed workflows, Docker Compose environments, tests, and CI/CD validation
• Applied ML Experiments — reinforcement learning, computer vision, model evaluation, and careful baseline comparison

A full-stack machine-learning web application for salary prediction. The system connects a React frontend, FastAPI backend, PostgreSQL database, model training pipeline, inference workflow, and containerized local development.

Engineering evidence

• Built separate training and inference responsibilities so model updates do not block the API service design.
• Compared multiple regression models and tuned candidates with Optuna.
• Reported model performance with R² = 0.899, MAE = 11,291, and RMSE = 16,611.
• Implemented Docker Compose local development plus GitHub Actions checks and GCP Cloud Run / Cloud Run Jobs deployment patterns.

Why it matters: this is my strongest end-to-end ML software project because it combines ML, backend APIs, frontend integration, database workflows, containers, CI/CD, and cloud-oriented architecture.

A local-first Retrieval-Augmented Generation tool for querying personal Markdown, text, PDF, and Python notes through a CLI / REPL workflow.

Engineering evidence

• Implemented ingestion, chunking, SentenceTransformer embeddings, and FAISS IndexFlatIP retrieval for local personal-notes scale.
• Designed citation-aware answers using source file path, document ID, and chunk ID.
• Added provider abstraction for OpenAI, Gemini, Ollama, and Hugging Face-style backends.
• Used JSONL metadata with byte-offset indexing for O(1) post-retrieval chunk metadata lookup.
• Reduced repeated index-update work by about 96% with changed-file smart re-indexing.
• Added tests, Docker support, CI/CD, runtime logs, and RAGAS-based evaluation.

Why it matters: this project shows practical LLM application engineering: retrieval design, grounding, configuration, observability, evaluation, and maintainable local-first AI tooling.

An experimental reinforcement-learning project for studying trading-policy behavior on historical market data. It is a learning and research system, not production trading software or financial advice.

Engineering / research evidence

• Implemented Q-learning, SARSA, DQN, and Deep SARSA agents.
• Compared epsilon-greedy and softmax exploration strategies.
• Designed a simplified buy/sell/hold environment with state, action, reward, and portfolio-value evaluation.
• Compared learned policies against a buy-and-hold baseline and saved plots, reports, Q-tables, and model weights.
• Best stored DQN epsilon-greedy historical backtest reached +855% return and 2.03x the buy-and-hold baseline under project assumptions.

Why it matters: this project demonstrates reinforcement-learning implementation, experiment design, baseline comparison, and awareness of evaluation limitations such as transaction costs, slippage, and historical backtest constraints.

I also completed postgraduate AI / computer-vision projects at UNSW. These are framed as academic project evidence rather than production systems.

• MRI Image Style Transformation — contributed to an academic group project comparing VGG19 neural style transfer with CycleGAN for unpaired Siemens-to-Philips MRI scanner-domain image transformation. The project demonstrated CycleGAN's reusable-generator workflow and measured generator-loss reduction, while avoiding claims of clinical validation.
• Solar Cell Defect Classification — contributed to a COMP9517 computer-vision project on ELPV solar-cell defect classification, comparing classical CV/ML methods and CNN architectures such as VGG11, ResNet-18, and MobileNetV3 under class imbalance.

I'm currently deepening my skills in:

• LLM evaluation — retrieval quality, grounding, faithfulness, and answer relevance
• Fine-tuning / steering fundamentals — efficient adaptation concepts and safe claim boundaries
• Agentic workflows — tool use, planning, memory, and multi-step AI-assisted engineering patterns
• MLOps — reproducible model workflows, packaging, deployment automation, monitoring, and CI/CD
• Cloud AI architecture — clean interfaces for deployable AI services and background ML jobs

Languages

AI / ML / LLM

Backend / Data / Product

DevOps / Cloud / Workflow

• Build projects so they can be run, tested, and explained, not just shown in screenshots.
• Prefer Dockerized, reproducible workflows and clear setup documentation.
• Use metrics and baselines, but keep claims scoped to what the project actually proves.
• Treat README files as engineering evidence: architecture, testing, limitations, and honest deployment status matter.

• LinkedIn: Yen-Jung Huang
• Email: yenjung178741@gmail.com