The AI and automation engineer at MomentScience, an ad-tech and rewards company, working directly with the founders, sales, and partners. I own the stack end to end.
Agent software factory running
Work arrives as noise. It leaves as a reviewed pull request. The cheapest thing that can do each step does it. How it works
Reads every item on the desk, keeps the actionable ones, links each to the task it belongs to. Zero cloud tokens.
A ranked inbox. Approve, dismiss with a reason, or snooze. Approved items become the runs below.
Email · Slack · transcripts · boards → local filter → your approval → a run
Run Read partner goal deadlines from the offer catalog
Classified locally: build. Not investigate, spec, or human. Zero cloud tokens spent.
Local vs cloud
Human in the loop · built-in testing
A gate is a durable state, not a prompt waiting on a keypress: the run parks for hours if it has to and resumes where it stopped.
Curated tasks in, reviewed draft pull requests out, with no prompt written per task. The run is animated above; the full design, including why there are exactly two human gates and the local-versus-cloud benchmark behind the model routing, is written up on its own page.
5,000 unstructured clinical notes turned into structured medical data: PySpark ingestion, Claude extraction of diagnoses, procedures, and medications, ICD-10 matching against the CMS 2025 code set, and a Streamlit review dashboard. Diagnosis F1 82%, procedures 56%, medications 48% against ground truth.
ClickHouse-backed revenue dashboards with hour-by-hour deltas and end-of-day projections, plus a Python bench that scores about 40 time-series models (SARIMA, Prophet, Chronos, TimesFM) with walk-forward CV across daily, monthly, quarterly, and annual horizons.
50+ live, branded prototype sites: each one wired to a production SDK so a prospect sees their own product with real offers inside it, provisioned on Render and Cloudflare with a single command.
A one-page operations dashboard that triages Gmail, Slack, Notion, and task boards into a single ranked action list with a locally served Gemma 4 model, backed by background gatherers so a dead source never blocks the page. Paired with a Claude Agent SDK Slack bot on a Hetzner VPS.
Production multi-tenant restaurant point-of-sale with order management, menu config, real-time reporting, and CI/CD, serving live restaurant clients.
Sole AI and automation engineer at an ad-tech and rewards company. Agent pipelines, customer-facing prototypes, a ClickHouse analytics and forecasting platform, partner integrations, and the production API behind a consumer rewards product.
Applied AI for clinical data (LLM extraction pipelines, ICD-10 validation, PySpark, Claude API, Databricks) and a multi-tenant cloud POS platform for live restaurant clients.
ML pipelines and data products for healthcare analytics, bridging data science and medical practice.
Intensive program: statistics, ML, NLP, distributed computing. End-to-end projects from acquisition through deployment.
Emergency and critical care across five facilities, including a Level II trauma center and Indian Health Services on the Cheyenne River reservation, through the COVID surge. Firsthand fluency in clinical workflows, EHR systems, and what documentation has to survive an audit.
Aerospace medical technician and instructor: trained 7,000+ personnel and certified 2,000+ in first aid and CPR. Before that, munitions systems: $1B+ in sensitive assets across three continents, and a SECRET clearance.
Open to fully remote roles in applied AI and forward-deployed engineering: embedded with customers, shipping agent systems that hold up in production.