AI Engineer · AI Systems Architect · Detroit, MI
Systems, not demos. Each one shipped, each one I built end to end, each with the shape of how it runs.
A company's AI infrastructure from zero: registration, training, attendance, retention, knowledge retrieval and natural-language access to business data. It replaced four spreadsheets and a paper sign-up process, company-wide.
A retrieval and agent framework for financial decisions, with telemetry, evaluation and monitoring in every step, fed by LSTM forecasting systems.
A system that listens to a live venue and adapts in real time: 100+ DSP features through PCA and KMeans into similarity scoring, with a continuous evaluation loop. The same signal pipeline reads sensor feeds and raw text: scrape, clean, extract features, then model.
Detroit work across one vehicle's whole life: how it is configured, how it is diagnosed, how it is serviced, and who services it honestly.
250,000+ vehicles and 108 option codes. Neural networks against logistic regression at three profit thresholds found 14 drivers, and one hard fact: about 95% of sales volume came from roughly 11% of configurations. The long tail was cost.
CVDOS vehicle diagnostics in Python, SQL, Spark, Hadoop and MATLAB: multi-source merges, automated daily reports in Jupyter, AWS SageMaker workflows. PCA and clustering grouped ECU faults into recurring defect patterns engineers could act on, and ECU issue resolution got about 20% faster.
About 1.47M work orders. Scheduling delay, rescheduling and lead time came out of the timestamps; make–model bins by cancellation ratio became risk tiers, so operations knew which vehicles to plan around.
About 500 automotive service shop locations. Random Forest, clustering and anomaly detection flagged the shops; SHAP showed the business why each one was flagged. About 20% more fraudulent locations found.
An always-on-top helper screen that runs my work: 20 Python packages, mail, usage, deadlines and a tray of agents. Each agent lives in its own folder, and a guard stops any reach outside it until I allow or deny it.
How every article on my company site is made: one model call writes it, Python laws check it (title, description, answer length, heading ladder, sources for every number), I read the test page and approve, and the same press pushes, deploys, pings IndexNow and checks Google's index through the Search Console API.
One headless Claude build per folder, jailed to that folder, with a cost cap and a snapshot first. Python gates then decide: files landed, everything parses, docs present, folder clean, tests ran. One repair at most, then a ticket, and undo is one command.
Years of single-purpose Python functions, each written and tested on its own, now run inside unified agents that take on whole jobs. I know what every step does, because I wrote every step.
The science under every system above: noisy inputs, honest models, numbers a decision can stand on.
250,000+ vehicles from Fiat Chrysler. Configuration codes became 108 binary features (engine, trim, interior packages); multilayer perceptrons in R, trained with SGD and binary cross-entropy, classified each configuration above or below $10k, $14k and $17k profit. Benchmarked against reduced multivariate logistic regression, the work narrowed 108 options to 14 drivers. The finding that moved production: about 95% of sales volume came from roughly 11% of configurations.
Gender differences in coronary plaque across 145 patients' CT angiography: plaque volume, low-attenuation plaque, plaque count, remodeling index, positive-remodeling probability and plaque type, controlled for diabetes, hypertension and smoking.
~1.47M unique work orders pulled from AWS S3 (boto3) into pandas. Time features engineered from appointment and opportunity timestamps (scheduling delay, rescheduling, lead time) showed how timing drives cancellations; vehicle make–model bins grouped by cancellation ratio became risk tiers.
One-to-one mentoring in statistics, machine learning, Python and R. Explaining a model to someone new is the fastest test of whether I understand it.