About
I'm a data engineer who builds the whole thing — pipeline, service, and the interface people actually use. Currently at Global X Investments, building Azure data platforms for a Canadian ETF manager; UofT MEng in Computer Engineering, with a BASc in Industrial Engineering and a Minor in Artificial Intelligence before that.
At Global X I architected an end-to-end Medallion platform on Azure SQL and Data Factory, consolidating fund-flow and AUM data from five systems into the model that powers sales reporting. Before that, 14 months at FedEx Express Canada on Azure Databricks and PySpark, replacing manual Excel reporting with automated Delta Lake pipelines. The through-line: take something manual and fragile, and make it run itself.
I design and ship my own systems. Dingyang Zhipai is a live judicial-auction property platform — five-source ingestion, cross-platform deduplication, and a two-stage LLM legal-review service, on React 19 Server Components. October is my always-on AI assistant, running across a local node and a cloud droplet with uncertainty-based model routing, a grounding pipeline, and an adversarial verification pass — plus a browser voice interface I built for it.
What connects both halves is a single conviction: LLMs belong behind deterministic, auditable controls. In production that means confidence thresholds that gate what gets written and what goes to a human; in my own systems it means the model interprets intent while retrieval, scoring, and ranking stay rule-based — same input, same output, every time. I care about the boundary between what a model decides and what a system guarantees.
Experience
Data Engineer Intern
GlobalX · Toronto · May 2026 – Dec 2026
Built the Azure data platform behind Global X's sales reporting: five scattered AUM and fund-flow feeds distilled into one governed Medallion model that Power BI — and the sales team — actually trust.
In one line: Five raw feeds in, one trusted revenue picture out — no manual steps left.
- Architected the end-to-end Medallion platform (Azure SQL + Data Factory) that consolidates five internal and external AUM/fund-flow sources into a single sales-territory fact model.
- Retired a fully manual ingestion-and-reconciliation routine: ADF-orchestrated pipelines and T-SQL procedures now run it end to end, holding <1% variance against validated outputs.
- Built an Azure Functions entity-resolution service — deterministic normalization ahead of fuzzy scoring — that auto-promotes ≥99%-confidence matches and routes the rest to human review, cutting manual matching ~70%.
- Gave the business its own eyes: structured run logs feed a self-serve Power BI outlier monitor, and dynamic row-level security scopes every region to its own territory.
Data / ETL Analyst
FedEx Express Canada · May 2023 – Jul 2024
Developed and maintained production ETL pipelines, data models, and BI dashboards supporting enterprise analytics across the Customer Experience and Global Trading Service teams. Familiar with the EDW and CE Datalake environments.
Scope: 14-month co-op · enterprise ETL, EDW & Power BI
- Developed Express Cases, Clearance Cases, and Quality Compliance Reports; maintained the LAC CES Report quarterly with the CE and GTS teams.
- Built downstream ETL pipelines in Databricks SQL transforming curated EDW data into analytics-ready tables, with null, uniqueness, and reconciliation validation controls.
- Designed structured data models with standardized join keys, conformed dimensions, and reusable semantic definitions for direct Power BI ingestion.
- Partnered with GTS, CE, and operational stakeholders to validate KPI definitions and deliver ETL outputs and dashboards in an Agile environment.
Career Service Coordinator
University of Toronto · Oct 2022 – Apr 2023
Supported student career development through data analysis, CRM management, and strategic workforce planning.
- Managed co-op job research and data categorization using CRM systems, ensuring accurate job tracking and seamless candidate placement.
- Conducted data-driven analysis on job market trends, identifying key employer preferences and optimizing student job search strategies.
- Developed interactive Power BI dashboards visualizing student employment metrics to improve career-service decision-making.
Product Manager Intern
Porsche HK & Macau · May 2022 – Jul 2022
Led data analytics and UX optimization initiatives for the Porsche China digital platform.
- Conducted marketing data analysis using Python's Pandas and NumPy, cleaning and transforming raw customer interaction data from the Porsche China website to extract meaningful insights.
- Designed predictive analytics models to assess user behaviour trends and conversion probabilities, improving website engagement by 15%.
- Developed interactive functional prototypes using Axure RP, refining website design to optimize customer navigation and conversion rates.
- Implemented sentiment analysis using LLMs and NLP techniques, processing customer feedback to identify high-impact UX improvements.
- Created data visualization dashboards in Power BI for real-time website performance monitoring.
Projects
DingYang SmartAuction (鼎阳智拍)
AI & Software · Full-Stack / AI Systems / Legal Tech
A foreclosure-property intelligence and collaboration platform: aggregates listings from five judicial auction platforms, deduplicates them into physical properties, vets legal risk through AI + lawyer review, and serves buyers with AI-understood, deterministically-ranked search — plus a shared review workbench for the internal legal/ops team. Currently covers residential foreclosures in Xi'an.
Tech: React 19, React Server Components, Cloudflare Workers, libSQL, Docker, LLM Pipelines, Recommender Systems
- Aggregates residential foreclosure listings from five judicial auction platforms (JD, Taobao/Ali, Gongpai, CBEX, CAA) through their legal data APIs, deduplicating cross-platform records into physical properties with full version history.
- Natural-language search where the LLM only extracts intent into whitelist-validated structured filters; retrieval, filtering, scoring, and ranking are fully deterministic — identical input always returns identical results.
- Two-stage AI legal-risk review (router model triage + deep-review model) feeding a deterministic, auditable release gate: un-captured announcements hard-block auto-release, and every listing needs lawyer approval or gate clearance before buyers see it.
- Internal lawyer workbench: four-lane review queue with claiming and automatic timeout release, optimistic locking, field-level verification, full audit logging, and a coordination inbox for buyer–lawyer appointments.
- Hand-rolled recommendation prototype combining EMA-decayed user interest profiles with correlation-based collaborative filtering over similar users' profiles — architected to swap in ALS matrix factorization once the user base grows.
Distilling LoFTR into Compact Student Networks
AI & Software · Deep Learning / Computer Vision / Model Compression · Winter 2026 · ECE 1508 Applied Deep Learning
Compressed the 12M-parameter LoFTR transformer into lightweight student networks (0.44M and 0.66M), reaching 18–27× parameter reduction at 86–152 FPS vs. the teacher's 52.5 FPS with 37% less GPU memory.
Tech: PyTorch, Knowledge Distillation, CNN, HPatches, Focal Loss
- Designed two lightweight student architectures (a CNN + shallow cross-attention hybrid and a pure dilated-CNN) achieving 18–27× parameter reduction while running far faster than the LoFTR teacher.
- Implemented a multi-level knowledge distillation framework combining coarse confidence distillation (KL divergence), intermediate feature alignment (MSE), and ground-truth supervision (focal loss).
- Recovered accuracy lost to KL-induced confidence shift through systematic threshold recalibration, reaching 52.7% AUC@10 and 93–96% recall, surpassing the non-distilled baseline (31.3%).
WellSight
AI & Software · Data Engineering / FinTech / ML
A Python-based financial data platform integrating the Plaid API to ingest and process raw transaction data into structured, analytics-ready datasets, with a modular RAG pipeline for intelligent financial querying.
Tech: Python, Plaid API, FAISS, Pinecone, RAG, GitHub Actions
- Built a Python-based financial data ingestion pipeline integrating the Plaid API, transforming raw transaction data into structured datasets for downstream analytics and retrieval workflows.
- Designed a modular RAG pipeline combining embedding-based retrieval, structure-aware chunking, and cross-encoder reranking to generate grounded responses, with semantic indexing using FAISS and Pinecone.
- Implemented automated CI/CD validation workflows using GitHub Actions — regression evaluations, smoke tests, and monitoring pipelines — so failures surface and recover quickly.
UniConn
AI & Software · Cloud Infrastructure / DevOps / Full-Stack
A cloud-native, AI-supported event management platform with a containerized microservices architecture on DigitalOcean, featuring a full observability stack, dynamic reverse proxying, and event-driven real-time updates.
Tech: Node.js, Docker Swarm, PostgreSQL, Traefik, Prometheus, Grafana
- Managed containerized decoupled services integrating backend, real-time, and vector retrieval pipelines via Express.js and Node.js microservices using REST and event-driven architecture, with SendGrid notifications and WebSocket updates.
- Architected and operated the DevOps infrastructure on DigitalOcean using Docker Swarm, preserving PostgreSQL state across container redeployments and restarts through Docker volumes.
- Designed stateless backend containers for horizontal scalability and implemented Traefik as a dynamic reverse proxy for automatic service discovery, routing, and configuration without manual intervention.
- Built a monitoring and alerting stack using Prometheus and Grafana integrated with DigitalOcean Droplet metrics to track CPU, memory, HTTP latency, and disk utilization, with automated email alerts for downtime events.
UConnect
AI & Software · Mobile Development / Social
A mobile application designed to enhance university student connections and campus life. Facilitates social networking, event discovery, and community building among students.
Tech: Mobile App, React Native, Node.js, MongoDB
- Student networking and profile matching
- Campus event discovery and recommendations
- Real-time messaging and notifications
- Interest-based community groups
Toronto Fire Truck Optimization
AI & Software · Operations Research / Analytics
An operations research project focused on optimizing fire truck deployment across Toronto. Uses advanced optimization algorithms and data analytics to improve emergency response times and resource allocation efficiency.
Tech: Python, Optimization, Data Analytics, Gurobi, GIS
- Facility location optimization models
- Emergency response time analysis
- Geospatial data processing and visualization
- Resource allocation recommendations
FCN: Fast Food Image Classification
AI & Software · Deep Learning / Computer Vision · Jan. 2023 – Apr. 2023 · Team Member
A CNN-based image classification project for multi-class fast-food recognition. Implemented and evaluated multiple neural network architectures including custom CNNs and the benchmark VGG-16 model using PyTorch.
Tech: Python, PyTorch, CNN, VGG-16, LaTeX
- Implemented multiple CNN architectures including custom designs and VGG-16 benchmark.
- Conducted systematic evaluation using confusion matrices, precision, and related metrics.
- Performed data preprocessing: image augmentation, normalization, and dataset splitting.
- Logged training curves and hyperparameters; authored full experimental reports in LaTeX.
IISE Analytics Capability Improvement
Business Analytics · Business Intelligence / CRM · Oct. 2024 – Mar. 2025
Lead Developer for a project supporting growth imperatives through advanced analytics and business intelligence solutions.
Tech: Power BI, MS Dynamics CRM, ETL, DMAIC
- Integrated MS Dynamics CRM with Power BI, building scalable data models for efficient reporting.
- Built interactive dashboards visualizing profitability trends, operational bottlenecks, and revenue insights.
- Designed ETL pipelines automating data ingestion and transformation for Lean Six Sigma decision-making.
- Applied DMAIC methodology improving data accuracy and supporting evidence-based decisions.
DaFlash Logistics Process Optimization
Business Analytics · Process Optimization / Analytics · Sep. 2024 – Dec. 2024
Student Researcher leading business process analysis and AI-driven optimization for transportation logistics workflows.
Tech: BPMN, Fishbone Analysis, AI Forecasting, Process Mining
- Conducted business process analysis using stakeholder mapping, BPMN modelling, and concept diagrams.
- Applied Fishbone and Pain-Gain analysis identifying 10+ key pain points affecting performance.
- Proposed AI-driven forecasting models for route optimization, cargo allocation, and service quality.
Quality Compliance Report at FedEx
Business Analytics · Data Engineering / Compliance · Oct. 2023 – Dec. 2023
Lead Developer for real-time quality compliance monitoring and automated reporting systems at FedEx Express Canada.
Tech: Power BI, Azure Data Studio, Databricks, ETL
- Designed and developed Power BI dashboards for real-time monitoring of quality compliance metrics.
- Built automated ETL pipelines with Azure Data Studio and Databricks ensuring data consistency and governance.
- Standardized compliance evaluation metrics, leading to a 15% reduction in reporting errors.
October: My Personal AI Long-Term Memory Assistant (Inspired by Iron Man's Jarvis)
OpenClaw Work · Agentic AI / Voice Interface / Personal Infrastructure · Ongoing · Personal infrastructure
An always-on personal assistant with persistent long-term memory, built on OpenClaw and governed by a hand-written operating contract: uncertainty-based model routing, hard factual-grounding rules, and adversarial verification. Reachable over Telegram, Discord, and WeChat, plus a self-built browser voice interface (MyVoice).
Tech: OpenClaw, LLM Routing, React, WebSocket, Redis, faster-whisper, Telegram, Discord
- Persistent local/cloud architecture: my machine handles everyday computer duty while a Toronto droplet runs the same chain 24/7 for scheduled digests — both reading and writing one shared long-term memory (local ⇄ cloud sync in progress).
- Uncertainty-based model routing: Terra (default, basic tasks) → Luna (conflicting sources or full derivations) → Sol (when the first two disagree or stakes are high) — routing keys on uncertainty, not task difficulty.
- Hard grounding contract (my main paper-reading tool): every factual claim is searched — Serper for general queries, Perplexity for finance and academic searches — and verified against the original page, never snippets alone, with sources attached.
- Adversarial verification: every definitive verdict must first survive an isolated second pass whose explicit job is to refute it.