I build AI agents that ship real software, and the tests that prove they work.
Bachelors in Artificial Intelligence from Purdue (2026). 2+ years building cloud, data-heavy systems on Google Cloud. Founding a local-first operating system for AI agents on my own time.
A working command line. Try one of these, or press / to focus it from anywhere.
I build AI systems, then I build the tests that prove whether they actually work.
BUILD stage 1 of 4 · design the system
Things I built, with the numbers they earned.
Open any case file. The first one is live: it pulls this week's forecast straight from the simulator.
Race intelligence · personal project
Flat Out F1
A lap-by-lap F1 race simulator that runs every Grand Prix 4,000,000 times: tyre wear, pit strategy, safety cars, overtakes and reliability. Every prediction is scored against the result in a walk-forward backtest.
4M
sims per race
60
Grands Prix audited
-12%
error vs grid baseline
Latest forecast loads when this card is on screen.
Python
NumPy
LightGBM
FastF1
GSAP
Playwright
Open case file
Pace model learns fuel- and tyre-corrected race pace from 60 audited Grands Prix (2024-2026), with calibrated uncertainty.
Predicts tyre strategy and pit windows, plus car-adjusted driver and constructor ratings.
Beats the "finish where you start" baseline by 12% on ranked probability score, retrained before every race.
An operating system for AI agents that puts your own machine first. Agent workloads run on-device by default, and when local resources run short it bursts to rented cloud GPU compute.
AI agents
Local-first
On-device compute
Cloud GPU bursting
Open case file
Local-first: agent workloads run on-device by default.
Bursts seamlessly to rented cloud GPU compute when local resources are constrained.
Temporal Fusion Transformer and TimesFM models forecasting supplier demand across 1,500+ suppliers, built under production schema and data-quality constraints.
13.4%
WAPE on held-out data
1,500+
suppliers
PyTorch Forecasting
Temporal Fusion Transformer
TimesFM
Python
pandas
Open case file
Owned the pipeline end to end: ingestion, feature construction, training, held-out validation and error analysis.
Integrated procurement and shipment data with strict no-future-leakage evaluation.
A high-throughput network traffic analysis platform with ML-based flow classification, supporting WAN capacity planning and anomaly detection for Caterpillar's IT org.
15M+
flow records per minute
Cloud Datastore
Python
Grafana
LDAP
Open case file
Built on Google Cloud Datastore with ML-based flow classification.
Grafana dashboards monitoring 15M+ network flow records per minute in real time.
LDAP-based authentication for role-aware access to analytics and WAN planning tools.
BREAK stage 2 of 4 · attack it before anyone else does
Then I try to break what I built.
benchmarks
Tests for frontier models
At Handshake I wrote software-engineering benchmark problems, reference solutions and automated judge tests that measure how well frontier models code, and scored model output for correctness, instruction-following, factuality and safety.
no future leakage
Forecasts that can't cheat
Supplier forecasts were evaluated on held-out data with strict no-future-leakage rules. The F1 simulator is scored against every real result in a walk-forward backtest, retrained before each race.
attack surface
Security from the inside
Intrusion detection with SIEM tooling on embedded hardware at Avolta, anomaly detection on Caterpillar's network traffic, and the Google Cybersecurity Professional Certificate.
MEASURE stage 3 of 4 · benchmark it against reality
The results, as a test run.
$ pytest career/ -v
✓test_supplier_forecast_wape13.4%TFT + TimesFM across 1,500+ suppliers, held-out data
✓test_network_flows_monitored15M+/minGrafana + Cloud Datastore, in real time
✓test_race_sims_per_grand_prix4,000,000Flat Out F1 Monte Carlo, live on f1.h
✓test_error_vs_grid_baseline-12%ranked probability score, walk-forward backtest
✓test_research_team_selection<10 of 100+applicants selected for the Purdue CS research team
✓test_years_on_google_cloud2+cloud, data-intensive systems
6 passed, 0 failed
I define how success gets measured early, then let those numbers drive the work.
SHIP stage 4 of 4 · put it in front of people
Where it shipped: eight roles, often two at once.
Each bar is a role. Overlapping bars ran at the same time.
Oct 2026 - PresentAI Engineer CurrentDevHawk AIAI agents for software delivery+
Build AI agents and LLM-powered tooling that take software work from idea to shipped code.
Aug 2026 - PresentFounder CurrentStealth startup · United StatesAgent OS local-first, cloud GPU burst+
Building a local-first operating system for AI agents that runs agent workloads on-device by default and bursts to rented cloud GPU compute when local resources are constrained.
Own the system architecture and implementation.
May 2026 - Oct 2026AI & Machine Learning Researcher (Contract)Handshake · RemoteFrontier model coding benchmarks+
Authored advanced software-engineering benchmark problems, reference solutions and automated judge test cases used to measure the coding capability of frontier models for leading AI labs.
Evaluated frontier LLM and multimodal model outputs against structured rubrics covering correctness, instruction-following, factuality and safety, and wrote evaluations used in downstream model assessment.
May 2026 - Aug 2026AI ResearcherPurdue University CS Dept. · West Lafayette, IN<10 of 100+ selected for the research team+
Selected as one of fewer than 10 students from 100+ applicants for an interdisciplinary research team advised by Dr. Qinglan Ding and Dr. Tianyi Zhang.
Designed and implemented the LLM orchestration layer of a personalized health chatbot for adults managing cardiovascular and metabolic risk: component interfaces, control flow and multi-step conversational workflows.
Architected a RAG pipeline with persona-aware context assembly that produces grounded, citation-traceable responses, and defined metrics for retrieval accuracy, grounding and unsupported-claim rate.
Aug 2025 - May 2026Data Science Researcher, Supply Chain ForecastingCaterpillar Inc. (The Data Mine) · West Lafayette, IN13.4% WAPE across 1,500+ suppliers+
Developed and fine-tuned Temporal Fusion Transformer and TimesFM models forecasting demand across 1,500+ suppliers, reaching 13.4% WAPE on held-out data.
Owned the end-to-end pipeline (ingestion, feature construction, training, held-out validation, error analysis), integrating procurement and shipment data under production schema and data-quality constraints.
Jan 2025 - May 2025Data Science Researcher, Network Security AnalyticsCaterpillar Inc. (The Data Mine) · West Lafayette, IN15M+/min network flows monitored+
Engineered a high-throughput network traffic analysis platform on Google Cloud Datastore with ML-based flow classification, supporting WAN capacity planning and anomaly detection.
Designed and deployed Grafana dashboards monitoring 15M+ network flow records per minute in real time, improving incident response.
Implemented LDAP-based authentication for secure, role-aware access to internal network analytics and WAN planning tools.
Aug 2024 - Dec 2024Data Science Researcher, Agricultural ForecastingBASF (The Data Mine) · West Lafayette, IN3 states corn + soybean forecasts+
Built geospatial ML models to predict corn and soybean production across Indiana, Iowa and Illinois.
Built scalable ETL pipelines integrating multi-source agricultural and remote-sensing data.
Dec 2023 - Jun 2024Machine Learning InternAvolta, Inc. · RemoteIDS + SIEM automotive intrusion detection+
Built Python software, ML models and intrusion detection systems (IDS) with SIEM tooling on Raspberry Pi / Raspbian hardware for automotive security.
Education and certificates
Bachelors in Artificial IntelligencePurdue University · 2026
Google AI Professional CertificateGoogle Career Certificates · May 2026
