Leon Chen

Mixed-integer program (Python + Gurobi) that rebalances inventory across a multi-sided supply network. Productionized with monitoring and scenario overrides.

Savings

$20M/yr

Scope

Multi-network

• Python
• Gurobi
• MIP
• Snowflake

Migrated UK/US retail forecasting from univariate ARIMA/ETS to a hybrid LightGBM + ensemble averaging the top-3 models per SKU.

WMAPE

−20%

SKUs

500+

• LightGBM
• ARIMA
• ETS
• Python

Org-wide agentic framework on Microsoft Copilot for data querying, analysis, and reporting. Deployed across 30+ internal projects.

Projects

30+

Surface

Org-wide

• Copilot
• LLM
• Agents
• Python

Difference-in-differences and synthetic-control methods to measure the incremental impact of optimization recommendations on service levels.

Methods

DiD · SCM

Audience

Leadership

• Causal
• Python
• R
• Statsmodels

Topic-driven personalized news digest. Scrapes RSS, Reddit, X, and podcasts, ranks via a four-pillar INRF formula (insight + relevance + niche) over pgvector embeddings, then synthesizes emails with a DeepSeek → OpenAI → Gemini → Grok LLM cascade. Next.js + FastAPI + Celery on Railway.

Stack

Next.js · FastAPI · Celery

Pipeline

4 stages, 4×/day

• Next.js
• FastAPI
• Supabase
• pgvector
• Celery
• LLM

AI bot trading arena where autonomous bots compete with $100K paper money on real US stocks (15-min delayed). Three FastAPI microservices on Railway share a Supabase Postgres, a polling order engine handles fills / margin / dividends / splits, and a React SPA on Vercel streams live prices over WebSocket.

Services

3 microservices

Seasons

2-month leagues

• FastAPI
• React
• Supabase
• Railway
• WebSocket
• X OAuth

Native iOS app (Swift + SwiftUI, MVVM) for real-time speech transcription, translation, and language-learning coaching. Uses the native Speech framework for continuous recognition and Google Gemini 2.0 Flash for on-the-fly translation and vocabulary extraction. Firebase Auth + Firestore history.

Languages

10 supported

Latency

Live captions

• Swift
• SwiftUI
• Speech
• Gemini
• Firebase

Two-stage virtual try-on: Google Vertex AI's try-on model generates the composite, then Gemini restores face, body, and garment detail before a 2K upscale. React + Vite frontend on Vercel, FastAPI backend on Railway, Supabase for auth, storage, and batch history.

Stages

Gen → Restore → 2K

Mode

N × M batch

• Vertex AI
• Gemini
• FastAPI
• React
• Supabase

Led the design and deployment of SKU-level demand forecasting models using XGBoost and Random Forest, improving forecast accuracy by 30%.

Accuracy

+30%

Grain

SKU-level

• XGBoost
• Random Forest
• Python

Product embeddings (Amazon Titan, CLIP, Doc2Vec) on large-scale product catalog to drive pricing and assortment strategy at Coach.

Signal

Visual + text

Use

Pricing · Assortment

• CLIP
• Titan
• Doc2Vec
• PyTorch

End-to-end retraining pipelines with Airflow + Kubernetes + Docker powering forecasting and pricing models at Coach.

Cadence

Auto-retrain

Infra

K8s + Airflow

• Airflow
• Kubernetes
• Docker
• AWS

Built ARIMAX time-series models in R and Python to forecast demand across Chinatex's textile accounts, incorporating exogenous drivers (promotions, seasonality, price).

Accuracy

+25%

Stack

R · Python

• ARIMAX
• R
• Python
• Time Series

Inventory management framework with dynamic safety-stock targets reacting to forecast error and lead-time variance — held the in-stock rate at 95% across the account portfolio.

In-stock

95%

Scope

Account portfolio

• Safety Stock
• Inventory
• Python
• SQL

Designed and automated ETL pipelines and Tableau dashboards for daily performance reporting, replacing manual Excel workflows and giving the commercial team same-day visibility.

Cadence

Daily

Audience

Commercial team

• ETL
• SQL
• Tableau
• Python

Flask web app on Heroku that builds a tailored route across 63 US national parks. A Python TSP solver returns itineraries in under 60 seconds, backed by an AWS RDS / SQLAlchemy schema and an ETL pipeline fed by the Google Maps and NPS APIs.

Satisfaction

9 / 10

Runtime

< 60s

• Python
• Flask
• TSP
• AWS RDS
• JavaScript

Multimodal classifier on the Hateful Memes benchmark using CLIP encoders fused via self-attention, cross-fusion, and MLP heads in PyTorch. Tuned learning rate, dropout, weight decay, and batch size against the challenge baselines.

Accuracy

+15%

AUROC

+14%

• CLIP
• PyTorch
• Attention
• Multimodal