I turn data into decisions —
and ideas into products.

A fueling and nutrition companion built for Type 1 diabetic cyclists — because I am one.

CrankFuel helps endurance cyclists who manage Type 1 Diabetes plan and execute their in-ride fueling: carbohydrate timing, insulin-on-board awareness, and glucose-aware nutrition strategies. It's the tool I always wished existed on long rides — built with safety as a first-class constraint, not an afterthought.

• Glucose-aware carb and fueling planning for endurance efforts
• Safety-first design — conservative defaults around insulin and lows
• Born from the intersection of two things I live: cycling and T1D

A complete refactor and modern rebuild of a law firm's web presence.

A ground-up rebuild of the Nicola Circelli Law website — modern, fast, and maintainable. Designed and shipped on a disciplined workflow: feature branches, pull requests, preview deploys, and a clean handoff. Proof that good engineering hygiene applies whether you're shipping to one client or millions.

• Full refactor to a modern, performant stack
• Branch → PR → preview-deploy → merge workflow throughout
• Backed by Supabase, deployed on Vercel

A unified lakehouse combining 60+ sources into one source of truth.

I architected Paystone's data foundation — originally a data lake on AWS S3 and Redshift unifying 60+ disparate sources, then consolidated into a Databricks Lakehouse for near real-time streaming and a far stronger MLOps experience. One governed platform that every downstream model, dashboard, and AI agent now runs on.

• 60+ data sources unified into one governed platform
• Consolidated onto a Databricks Lakehouse for near real-time streaming & MLOps
• Unified data and KPIs across the entire organization
• The foundation every model, dashboard, and agent is built on

Ask the data questions in plain English — no SQL, no waiting on an analyst.

Deployed Databricks Genie on top of the lakehouse to put a natural-language interface on company data. Teams ask questions conversationally and get governed answers back in seconds — self-serve insight at the speed of a sentence, instead of a dashboard request and a two-day wait.

• Natural-language querying over the Databricks Lakehouse
• Self-serve answers without writing SQL or waiting on analysts
• Conversational AI brought to everyday business decisions

Production AI agents — from pre-sales research to RAG-powered support.

Designer and builder of Paystone's agentic AI frameworks, with two agents in production. A pre-sales research agent profiles a prospect end-to-end before every call, saving reps roughly an hour each. A technical knowledge-base agent does retrieval over Databricks Vector Search (RAG) and is served flexibly — via MCP, a Databricks App, or directly inside Salesforce chat.

• Pre-sales research agent — full prospect discovery, ~1 hour saved per call
• Technical knowledge-base retrieval (RAG) on Databricks Vector Search
• Served anywhere: MCP, a Databricks App, or Salesforce chat
• Core to Paystone's AI-first transformation

An XGBoost model predicting which customers are likely to leave a Google review.

Built a gradient-boosted (XGBoost) model that scores each customer's probability of leaving a Google review. It lets the business concentrate review-generation effort exactly where it's most likely to convert — turning behavioral signals into measurable online-reputation growth.

• XGBoost model scoring review-likelihood per customer
• Focuses review-generation effort where it actually converts
• Turns behavioral signals into reputation growth

One reporting layer to consolidate every metric across the company.

Rolled out Looker as Paystone's company-wide reporting standard — consolidating metrics that had been scattered across tools and teams into one governed layer. A single, trusted definition of every KPI, consistent across the whole organization.

• Company-wide reporting standard on Looker
• Consolidated metrics across every team into one layer
• Governed, consistent KPI definitions org-wide

An internal tool that saves hours of manual candidate screening.

Built an internal resume-screening tool to take the grind out of shortlisting candidates — saving the hiring team hours per role while keeping a human in the loop for the decisions that matter. A small, practical example of pointing ML at the boring problems.

• Automated first-pass candidate screening
• Hours saved per role, human-in-the-loop by design
• Practical ML applied to real operational friction