scroll

Me

A little about me

Hello!
My name is Eddie, and I'm a student at the University of British Columbia pursuing a BSc in Computer Science with an option in Artificial Intelligence. I'm passionate about deep learning, especially generative AI and computer vision, and I aspire to become an AI research engineer. Right now, I'm tuning the retrieval in my RAG system, Zoned, and running my own PyTorch experiments.

In my free time, I enjoy playing badminton, watching shows, and eating good food.


Technical Skills

Languages

  • Python
  • Java
  • SQL
  • JavaScript
  • TypeScript
  • HTML/CSS
  • C/C++
  • R

Frameworks

  • PyTorch
  • FastAPI
  • Spring Boot
  • React
  • Streamlit

Libraries

  • scikit-learn
  • torchvision
  • NumPy
  • pandas
  • Matplotlib
  • Docling
  • psycopg
  • Alembic
  • Rasterio
  • Deck.gl
  • pytest
  • JUnit

Databases

  • PostgreSQL
  • pgvector
  • MySQL
  • Supabase

Tools & APIs

  • Git
  • Docker
  • Linux
  • Maven
  • REST APIs
  • Claude API
  • OpenAI API
  • Gemini API
  • Railway

Projects

Mini Eddo Logo

CFPB Complaint Triage

Multi-Agent RAG System · 10/2026 – 11/2026

Repo

A multi-agent system that drafts cited, compliance-safe replies to consumer finance complaints. It searches 5,163 real CFPB complaints and 413 sections of federal Regulations E, F and Z.

  • Two LangGraph workflows: a complaint-reply pipeline that waits for human approval, and a supervisor-routed analyst agent. Eight stopping conditions, enforced in code rather than the prompt, keep every run bounded.
  • Layered hallucination controls: a retrieval score floor, per-claim citations, code checks, an LLM fact-checker and human approval. Zero invented rules across 153 scored runs: every failure was an incomplete answer or a hand-off to a person.
  • Measured improvement: raised the pass rate from 61% to 93% on 28 test cases (3 runs each), then scored 90% on 23 new cases it was never tuned on.
  • Cost-aware design: about $0.05 per request, by routing classification and planning to a small model with 20× cheaper input. Three times the context cost 40% more with no accuracy gain, so lean context stays the default.
  • Python
  • LangGraph
  • Claude API
  • pgvector
  • PostgreSQL
  • streamlit
Mini Eddo Logo

Zoned

Citation-Grounded Document Q&A · 07/2026 – Present

Repo

A RAG pipeline that answers questions only from its source documents. Every answer cites the document and page, and it explicitly refuses when the corpus doesn't contain the answer.

  • Pipeline: ingests PDFs with Docling, embeds via the OpenAI API, and generates answers with Claude.
  • Evaluation on 33 questions: 23/23 document-level recall, 19/23 top-1 attribution, 10/10 correct refusals and zero fabricated answers. Every ranking failure traced back to nearly identical document pairs.
  • Single-database design: one PostgreSQL + pgvector schema for documents, chunks and 1536-dim embeddings. Retrieval is raw SQL over psycopg 3, with an HNSW cosine index, a generated tsvector column and Alembic migrations.
  • Python
  • PostgreSQL
  • pgvector
  • Docling
  • OpenAI API
  • Claude API
  • Docker
Mini Eddo Logo

ML Experiments

Self-Directed Deep Learning Research · 10/2026 – Present

Repo

A course assignment extended into a small research question: what's the best way to adapt an ImageNet-pretrained ResNet-18 to 10 PASCAL superclasses?

  • Compared two methods: averaging subclass head weights vs. summing subclass softmax probabilities.
  • Result: probability summing won both zero-shot (95.6% vs. 92.4%) and after head-only fine-tuning (96.9% vs. 95.3%), consistently across all 3 seeds.
  • A twist: adding an "others" class for the 774 unmapped ImageNet classes reversed the result (88.9% vs. 95.3%).
  • Python
  • PyTorch
  • torchvision
Momentum Logo

Momentum

Task Planner with Natural-Language Parsing · 04/2026 – 09/2026

Live Demo

Repo

A full-stack task planner where you create, edit, categorize and track tasks by deadline, deployed live on Railway behind a Docker build.

  • Full-stack build: Spring Boot REST backend with a React frontend.
  • Natural-language task entry with Claude: a forced tool schema for validated output and a confirmation step before any database write. If the API is unavailable, it falls back to raw input.
  • Database design: a normalized MySQL schema makes the next task a direct indexed lookup (verified with query plans). Foreign keys prevent orphaned tasks, and transactions make reordering all-or-nothing.
  • Java
  • Spring Boot
  • React
  • MySQL
  • Claude API
  • Docker
  • Railway
ReLeaf Logo

ReLeaf

Hack the Coast · 1st Place & Best Use of Gemini API · 02/2026

Live Demo Backend API

Demo Video

Repo

A 3D urban-heat planner that maps Vancouver surface temperatures. It projects the cooling, CO2 offset and ROI of placing trees, cool roofs and bio-swales.

  • Geospatial backend: FastAPI + Rasterio for per-pixel temperature lookups and hotspot ranking, validating placements against OpenStreetMap data.
  • Gemini integration: renders proposed interventions into real Street View images (cached in Supabase) and drafts FEMA grant narratives from the simulation results.
  • Python
  • FastAPI
  • NumPy
  • Rasterio
  • React
  • Deck.gl
  • Supabase
  • Gemini API
  • Google Map Tiles API
Portfolio Logo

Digital Portfolio

Portfolio built from scratch

Repo

This site: a responsive personal portfolio built from scratch in HTML, CSS and JavaScript, with no frameworks.

  • Animated hero: a CSS-only moon with radial orbit navigation.
  • Scroll reveals: driven by IntersectionObserver.
  • Accessibility: full keyboard navigation and reduced-motion support.
  • Deployment: hosted on GitHub Pages.
  • HTML
  • JavaScript
  • CSS
  • GitHub Pages
Mini Eddo Logo

Folder Sorter

Academic Project

A hierarchical file management system that creates and sorts nested folder/file relationships.

  • Persistence: JSON-based saving and loading.
  • Interface: a GUI built with Java Swing.
  • Testing: JUnit tests to keep the code stable.
  • Java
  • Java Swing
  • JUnit

Contact

Download PDF
Eddie Yang's resume