I build practical AI workflows, not proofs of concept. These projects connect real tools, solve real problems, and produce outputs that people actually use. Each one started as a question: what would happen if I automated this? The answers are below.
Project 1
Velocity – AI Content Pipeline
Velocity turns a two-minute audio clip into four publication-ready content formats: a short LinkedIn post, a long LinkedIn article, an SEO blog post, and a newsletter paragraph. Built for executives and marketing teams who have the ideas but not the time to write across channels.
The pipeline transcribes audio using Whisper, enriches the transcript with real-time research via Tavily, then passes everything to Claude with a custom style guide and content library to generate output that sounds like the person — not like AI.
Stack: n8n · Claude API · Tavily · Whisper · HubSpot
Demo coming soon.
Project 2
Email Segmentation Pipeline
An n8n workflow that takes a contact list, uses Claude to segment the audience by role, industry, or pain point, generates personalized email variations for each segment, and delivers them through HubSpot. Built to move beyond broadcast email toward genuinely targeted outreach at scale.
Stack: n8n · Claude API · HubSpot
In development.
Project 3
Material Specs RAG Pipeline
A Python RAG system built to answer precise technical questions across 33 industrial manufacturer PDFs — adhesive specifications, application guidelines, and performance data. Users query in plain English and get exact values with source citations. The system doesn’t guess.
Solved two specific retrieval problems: increased chunk size from 500 to 1,000 tokens with 200-token overlap to preserve table context, and implemented query expansion to handle terminology variation across manufacturers.
Stack: Python · PyMuPDF · OpenAI embeddings · Pinecone · Flask · Render · GitHub
[View on GitHub] [Live demo] [Read the case study]
Project 4
Material Specs KB – No-Code RAG Assistant (Flowise)
A no-code RAG pipeline built on 30 industrial manufacturer PDFs using Flowise, OpenAI, and Pinecone. Non-technical staff query material specifications — heat resistance, viscosity, tensile strength — in plain English and get exact values with source citations.
Designed for maintainability: adding a new data sheet requires nothing more than dropping a file into a Google Drive folder and clicking re-index. No technical intervention needed.
Stress tested for accuracy, cross-document retrieval, and hallucination resistance. The system declines to answer when information isn’t in the knowledge base — which in a materials specification context matters as much as what it does answer.
Stack: Flowise · OpenAI · Pinecone · Google Drive
Project 5
ServiceNow NowAssist – AI Agent Optimization
At illumin, the AI support agent was returning poor answers. The problem wasn’t the AI — it was the content underneath it. Search failures and test queries pointed to structural problems: weak answer-first formatting, poor chunking, taxonomy mismatches across three platform versions.
I restructured the KB architecture. The agent started surfacing accurate answers instead of failing silently. This project shaped how I think about AI-ready documentation: the quality of the output is only as good as the quality of the content the AI is trained on.
Project 6
AI Learning System – ESL Listening Feedback
A mobile-first AI feedback system built inside a LearnPress course. Learners respond to listening prompts and receive immediate meaning-focused feedback from Claude — not grades, not right/wrong scores, but coaching that points out what they understood and what they missed.
The AI follows strict instructional rules: check idea coverage not grammar, keep feedback brief for mobile reading, encourage retries rather than final answers. Built with WordPress, LearnPress, and server-side OpenAI API calls.
More projects in progress. This page is updated as work ships.