Portrait of Eric Wang

Eric Wang

I’m a data science student at the University of Toronto. My work spans LLM applications, information retrieval, machine learning, and statistical modelling.

  • Study
  • Research
  • Work
  • Projects
  • Activities

Bars show durations and dots mark the month a project was completed. Bars that reach the October 2026 line are ongoing; the dashed segment is expected.

About

I want to work on AI in practice: taking existing models, such as large language models, image models, and classic machine learning, and turning them into working applications that get real tasks done. I want data science and AI in practice to work as two halves of one project. Data science decides what is worth building, prepares the data, and turns the results into decisions. AI in practice builds the system, ships it, and monitors it. The two meet at evaluation, where the system is tested before anyone relies on it. I’m exploring healthcare, public health, and the environment as places to apply this, and I’m open to others.

Education

Honours Bachelor of Science

University of Toronto (U of T)

– (Expected)

Toronto, Canada

Data Science Specialist, Statistics Major, and Computer Science Minor

Relevant courses: Machine Learning, Deep Learning, Data Science, Databases, Time Series Analysis, Probability & Statistics, Data Structures & Analysis, Software Design, Methods of Data Analysis, Survey Sampling, Advanced Calculus, Software Tools & Systems Programming

Research

Correlation-Augmented Neural Retrieval

U of T

– Present

Toronto, Canada

  • Research project, supervised by Prof. Nick Koudas (Department of Computer Science), testing whether word-association statistics (a term-association graph from CRAFT, Corpus Relatedness Analysis Using Fourier Transforms) can help AI search (dense retrieval) cut results that miss a query’s topic and find passages that answer it without sharing its words; so far built the baselines and search indexes the experiments will use.
  • Evaluated and optimized information retrieval configurations, measuring embedding throughput, index size, CRAFT pipeline time and memory, and fine-tuning (LoRA) feasibility, enabling efficient experimentation on a laptop.
  • Matched the model cards’ search-quality scores (nDCG@10) within ±0.01 on all 4 baseline information retrieval runs (text embeddings, exact search), generating BGE embeddings for 4 corpora and E5 embeddings for 2.
  • Met the model evaluation target of 0.95 recall@100 on all 6 development indexes, building 8 vector-search indexes (HNSW in Faiss and hnswlib) and a search implementation whose result-matching checks returned the same top-100 results as Faiss on every development query.
  • One of three undergraduate students on the project, which is in progress.

Experience

Data Operations Engineer (Co-op)

AISHU Technology Corp.

–

Shanghai, China

  • Developed LLM applications (Text-to-SQL, Chat BI) that enabled users to question 3 customer databases in natural language, with data engineering (writing and optimizing SQL); answers took about 1–13 minutes for questions that could not practically be answered by hand.
  • Designed about 400 SQL test queries for the model evaluation (execution accuracy) of the first 2 proofs of concept (POCs), running 9 rounds of business-rule tuning and prompt engineering (DeepSeek V3) across all 3; those 2 POCs each improved from about 50% to about 70% execution accuracy.
  • Turned customer business rules into 30+ Markdown documents (about 6,000+ rows after chunking, across 3 POCs) for the LLM’s context engineering, with stakeholder management across 5 customer consultations and follow-up, and personally demonstrating to customers the 2 POCs that were delivered.
  • Held majority responsibility in a 3-person team for designing the SQL evaluation queries and customer business rules.

Data Analyst & Logistics Assistant

Marubeni-Itochu Steel Canada Inc.

–

Toronto, Canada

  • Produced an Excel report for a trend analysis of 5 years of steel imports into Canada, with data cleansing and data analysis of Statistics Canada data comparing net weights before, during, and after COVID across 10 product categories and 5+ partner countries, and identifying the shifts the analysis used.
  • Created data visualization (charts) and executive summaries for internal stakeholders as part of stakeholder management, maintaining regulatory-change datasets that supported planning and reporting.

Projects

My Second Brain by YFW

Personal project

– Present

Python, TypeScript

Code for My Second Brain

  • AI knowledge app that runs locally and turns text, documents, images and scanned PDFs (vision transcription), and web pages into a wiki for Obsidian that links each page to its sources, so answers can be traced, and reconciles changed or deleted Markdown incrementally; v1.0.0 passed all 7 citation-grounding checks.
  • Delivered hybrid information retrieval for answers that cite their sources (RAG), fusing keyword (BM25) and vector (LanceDB) results by Reciprocal Rank Fusion within each knowledge scope, adding LLM reranking as an option, and falling back to keyword search when vector search fails.
  • Deployed v1.0.0 as a Python package, command-line tool, and web app (FastAPI, TypeScript, Vite, Pydantic) with OpenAI and Anthropic adapters, passing 176 Python tests in its software engineering process.

Global CO2 Per-Capita Emissions Analysis with World Bank API

Independent project

–

Python, Quarto

Live site for the CO2 emissions analysis

Code for the CO2 emissions analysis

  • Data science project identifying which GDP, energy, forest, and urbanization indicators best predict per-capita CO2 emissions in 196 countries (1990–2023); XGBoost gain-based importance ranked log GDP, renewable-energy share, and urbanization as the three highest-importance features.
  • Built a 7,378-row country-year panel in a pandas data pipeline, with data engineering (6 World Bank API indicators combined by ISO3 joins), data cleansing (region filtering), and feature engineering (log transforms, year centring), keeping 5,499 complete country-year cases.
  • Reached a test R² of 0.722, with error metrics (RMSE 0.543, MAE 0.359), from an XGBoost machine learning model under leakage-resistant model evaluation (70/15/15 country split, country-grouped 5-fold tuning), comparing it with regression models (OLS, GAM) and a pruned decision tree (R² 0.706).
  • Published the results’ data visualization as a reproducible Quarto/GitHub Pages site with 3 Plotly figures, including map animations, distributions by decade, and GDP-emissions views.

LLM Source Classification with Survey and Text Features

Team project

–

Python

Code for LLM Source Classification

  • Machine learning project identifying whether ChatGPT, Claude, or Gemini wrote a response, from 825 survey and free-text observations; a feed-forward neural network led on the held-out test set with 69% accuracy (0.69 macro F1).
  • Inspected the survey records and, with data cleansing, standardized labels and encoded Likert and multi-select answers.
  • Constructed the feature matrix, combining the survey features with the project’s text features (TF–IDF, vocabulary cap) in feature engineering and preprocessing train, validation, and test splits the same way.

NLP Classification: Tweet Sentiment and Text Emotion

Team project

–

Python

  • Text classification of COVID-19 tweet sentiment (3 classes) and the emotion of 18,000 text samples (6 classes); the emotion model chosen in tuning reached 84% test accuracy.
  • Compared Naive Bayes and softmax regression with NLTK feature engineering (word counts, TF–IDF) for NLP.
  • Selected a 2,000-component emotion model (word counts, dimension reduction by PCA, softmax regression) after model evaluation with 5-fold stratified cross-validation (83.5% mean validation accuracy), tuning the component count.

Global Life Expectancy Regression Analysis

Team project

–

R

  • Studied how health, economy, education, and development indicators relate to life expectancy (WHO dataset, 1,649 observations), identifying schooling, GDP, adult mortality, and income composition of resources as predictors.
  • Refined the project’s statistical modelling in R: multiple linear regression with variable transformations, multicollinearity checks (VIF), partial F-tests, and model selection (AIC).
  • Validated model performance and made results interpretable, using regression diagnostics and data visualization (ggplot2) to ensure statistical assumptions were met.

Meal Master

Five-person team

–

Java

Code for Meal Master

  • Java app that allows users to search, save, and organize recipes, manage accounts, and generate shopping lists (TheMealDB and File.io APIs), with a modular architecture for maintainability and testability; proposed its concept.
  • Implemented the core of the user and recipe data-access objects with modular software engineering design, taking responsibility for the Shopping List use case.
  • Reviewed code in its software engineering process for correctness, tests, architecture, style, and documentation.

Activities

Volunteer Note-Taker, Financial Accounting

U of T Accessibility Services

–

Toronto, Canada

  • Took Financial Accounting notes for students using accessibility services and uploaded them within 48 hours after each lecture.

Member, Recognized Study Group (Calculus with Proofs)

U of T

–

Toronto, Canada

  • Met at least once a week to discuss challenging class or homework problems, and completed a reflection to improve learning and group communication.

Skills

Core skills
Data Science, Data Analysis, Data Cleansing, Data Pipelines, Data Engineering, Statistical Modelling, Machine Learning, Model Evaluation, Natural Language Processing (NLP), LLM Applications, Information Retrieval, Data Visualization, Software Engineering
Software and programming
Python, SQL, R, C, Java, Git, Quarto, MS Office Suite
Data science
pandas, NumPy, SciPy, scikit-learn, XGBoost, TensorFlow, statsmodels, pygam, PyTorch, Hugging Face Transformers, Sentence Transformers, Model2Vec, Faiss, hnswlib, NLTK, LangChain
Visualization
Plotly, Folium, plotnine, Tableau, Matplotlib
AI engineering principles
Prompt Engineering, Context Engineering, Harness Engineering
Coding agents
Claude Code, Codex, Cursor
Languages
English, Mandarin
Interests
Swimming, piano, Chinese history, and watching football (soccer)

Certifications

DeepLearning.AI and Stanford University
Supervised Machine Learning: Regression and Classification
DeepLearning.AI
Generative AI for Everyone
LangChain for LLM Application Development
Building Systems with the ChatGPT API

Contact

Email is the best way to reach me.

Email
yifanwang02@outlook.com
LinkedIn
Eric Wang on LinkedIn
GitHub
github.com/EricWcr7
CV
Open my CV (PDF)
Location
Toronto, ON, Canada