Eleven-plus years shipping production data and machine-learning platforms in Fortune 100 environments. Today I build the pricing models behind Sherwin-Williams. After hours I run a lab of agents where every output is graded against a baseline before it’s allowed to matter.
I take a business problem, build the system that answers it, and prove it works before anyone relies on it.
I started in data engineering and moved steadily toward machine learning, then toward agents. At Sherwin-Williams I own the full lifecycle of the pricing-structure models. Before that I led a 25-person AI/ML organization at T-Mobile, and earlier built payments, fraud and compliance systems at Global Payments and Compliance.AI.
Outside work I build and run a lab of read-only agents that collect market data, detect signals and report by email. It exists to practice a discipline I care about: evaluation harnesses, guardrails, permission boundaries, kill switches and monitoring, written and debugged by hand.
Most of what it measures turns out to have no durable edge. That’s exactly what the harness is for.
Every output is graded against a baseline, net of costs. Only what clears the test is surfaced; the rest is logged and withheld.
Read-only scopes, confirm-before-action rules, credentials kept out of prompts and logs, and kill switches when live results decay.
Feature engineering through training, validation, deployment and monitoring, then supporting adoption after release.
Applied ML and data platforms, from startup NLP to enterprise pricing and telecom operations. Open any role for the details.
Build and operate the ML models behind the pricing structure, working directly with business stakeholders and external partners.
Led AI/ML platform engineering and delivery for telecom operations.
Generative AI, MLOps and responsible AI on payment-scale data platforms.
Master-data hub for 15 global manufacturing sites.
TensorFlow and spaCy models for regulatory-document classification, with Airflow/Docker pipelines and REST APIs.
Read-only agents that collect data, detect signals and report by email. None of them places orders. Each card says what the evidence actually showed.
A shared 15-minute market-data store and a signal ledger every agent writes to. Each signal is graded net of costs, tagged with its market regime and compared with random entry. A signal kind is only “proven” after 30+ graded signals over 10+ days with a Wilson lower bound above baseline; until then it stays out of alerts.
Read-only agents that call Kalshi’s 15-minute up/down contracts only when the price cushion is large against the volatility still to come. The gate is set by a backtest rule fixed in advance and checked by a live scorecard. The gold and WTI version adapts the model to a single Pyth settlement candle and a proxy feed whose error is measured.
A Databricks notebook comparing an internal demand forecast with a vendor forecast: WMAPE, bias and win rates by scope, customer, brand and demand class, with data-quality guards on every join.
A headless watcher over 75 symbols that detects trend flips and early rallies on 15-minute bars, and emails only the signal kinds the evidence ledger has proven.
A read-only options-volatility scanner across about 517 names by sector. It proposes bull call spreads, and each is graded against the market’s own odds at expiry.
Watches new Solana and Base coins and Coinbase listings, with honeypot filters, a held-out hit-rate gate, a kill switch and a sentiment layer that is tested rather than assumed to predict.
Chronos-2, Chronos-Bolt and AutoGluon forecasts of MNQ and MES daily excursions over three years of bars.
A daily ZIP-level digest for 51 Georgia ZIP codes, built with zero dependencies. It works around a captcha-blocked source by tiling the map and splitting dense areas adaptively.
A Smart Money Concepts toolkit: a TradingView Pine indicator, a Python port with an R-multiple backtest, and a read-only scanner.
A shared client for an external LLM judgement service, run in shadow mode: every judgement is stored beside the outcome and graded before it can influence a decision.
No projects in this category.
Research and engineering projects, not investment advice. The agents are read-only and hold no credentials.
Pulled from my public repositories automatically every few hours, so this section, the project cards and the repo count stay current on their own.
What I reach for day to day, grouped by the kind of work.
An actuarial-science start, a technology degree, and the cloud and ML certifications that followed.
Open to conversations about applied ML, forecasting and agent systems. Email is the fastest way to reach me.