Portfolio
AI Engineer Data Scientist Software Engineer
Jack Lutz
I build software, models, and tools that turn messy problems into working systems.
Tech Stack
Skills and Technologies
Language Python
Models, data pipelines, APIs, and automation.
Experience
Researcher
UCLA Department of Mathematics
- Researching statistical arbitrage in cryptocurrency markets: constructing normalized spread series across liquid trading pairs and testing for mean reversion via cointegration analysis with walk-forward evaluation.
- Built a survivorship-bias-free dynamic universe spanning 100+ liquid spot pairs, rebalancing daily on volume and spread thresholds.
- Tested mean-reversion hypotheses via Engle-Granger and Johansen cointegration on rolling windows; quantified half-life of mean reversion to size positions.
Founder
Lutz Consulting Group
- Analyzed 138 months of client transaction data; built Gradient Boosting forecasting model with engineered lag, rolling, and seasonal features, achieving 11.5% MAPE, a 58% improvement over naive baseline on 24-month holdout, uncovering $550K+ in dead stock for liquidation.
- Engineered computer vision pipeline using GPT-4 Vision and AWS S3 to extract and classify 40,000+ product tags; built Streamlit review dashboard backed by PostgreSQL, enabling staff to reconcile full inventory in one weekend.
- Shipped end-to-end ML pipelines (data ingest → model → review UI → action) for non-technical clients; owned scoping, modeling, validation, deployment, and training.
Data Analyst Intern
New York Life
- Built RAG-enabled underwriting chatbot prototype using company underwriting manual as knowledge base, demonstrating feasibility of LLM-powered policy guidance.
- Designed controlled experiments measuring LLM sycophancy: varied prompt authority framing across multiple models, measured response drift toward user-implied conclusions, and produced the team's first empirical model behavior finding that drove production prompt engineering policy.
- Finalist in company-wide AI Innovation Competition for prototyping RAG-enabled chatbot plugin for Tableau dashboards, querying 10+ years of nationwide data with office-level granularity.
Projects
Deploying Real Systems.
Research platform
Research platform
Candella Quant
A full-stack research workspace that brings strategy development, backtesting, live charting, and AI-assisted analysis into one product.Features
- Co-developed C++ engine with a custom strategy language and walk-forward validation
- Research stack spans market data, strategy execution, and 27 MCP tools
Research platformCandella Quant
A full-stack research workspace that brings strategy development, backtesting, live charting, and AI-assisted analysis into one product.
- Co-developed C++ engine with a custom strategy language and walk-forward validation
- Research stack spans market data, strategy execution, and 27 MCP tools
TypeScriptPythonC++PostgreSQL
Agent benchmarkWikipedia Speedrun Benchmark
An interactive benchmark for comparing how AI agents navigate between Wikipedia articles using only the links available on each page.
- 420+ games across 35 routes and 15 agent strategies
- Embedding agents reached a 94% win rate at zero inference cost
PythonHugging FaceDockerEmbeddings
ML labLearning ML
A growing lab of from-scratch machine learning projects, beginning with scalar autograd and neural networks built to understand every training step.
- Autograd engine and neural-network library built from scalar operations
- Animated training diagnostics with gradients verified against PyTorch and JAX
PythonNumPyPyTorchmatplotlib
Menu dataConstraint solverMeal plan
Recommendation systemUCLA Dining Recommendation System
A data pipeline and constraint-based recommender that turns UCLA dining menus into meal plans built around each student's macro targets.
- Go ingestion pipeline loads daily menu data into PostgreSQL
- OR-Tools solver composes meal plans around nutrition and dining constraints
GoPostgreSQLPythonFastAPI
Private product
PyTorch DDPtinynccllibibverbs
Systems infrastructuretinynccl
A from-scratch C++ collective-communication library that routes PyTorch distributed training through libibverbs over softRoCE.
- Custom c10d backend sends DistributedDataParallel gradient sync through the library
- Trained a 10.7M-parameter character GPT across two ranks on TinyShakespeare
C++CUDAlibibverbsPyTorch
Native utilityClaude-o-Meter
A native macOS menu-bar and Windows tray app for monitoring real Claude and Codex subscription usage windows without opening another dashboard.
- Live 5-hour and weekly windows sourced from the providers' usage data
- Swift on macOS and Rust with Tauri on Windows, both released through CI
SwiftRust
ContactLas Vegas / Los Angeles
Have a difficult problem or an ambitious build?
Let’s Make Something.
I’m always interested in software, applied ML, data, and thoughtful product work.
Start a conversation