Designed. Built. Shipped.

The problem: A 600-person architecture firm was developing a new predesign workshop process. Design teams needed a way to ground early-phase sustainability strategy decisions in real precedent — but that data was scattered across 25 years of award submissions, PDFs, and institutional memory.

What I built: A 284-project database with 324 data columns and 211 boolean strategy dimensions, scraped and enriched from multiple sources (web archives, Obsidian vaults, PDFs). The matching engine lets teams upload their strategy selections and returns ranked precedent projects with multi-factor scoring. It's now integrated into the firm's predesign workflow and used by design leads to back strategy decisions with real data.

Why it matters: This tool replaced a manual process that would have taken a team of four working full-time for over a year. It connects predesign strategy workshops directly to evidence-based precedent, changing how the firm approaches early-phase design decisions.

The problem: 64 instructional dance videos contained hundreds of techniques, drills, and concepts — but finding a specific technique meant scrubbing through hours of footage. The knowledge was locked inside the video format.

What I built: A full RAG pipeline. Videos were transcribed, split into 2,500+ semantic chunks, and embedded using Sentence-Transformers (all-MiniLM-L6-v2) into a local vector store. The search engine finds relevant chunks via cosine similarity, then Claude API synthesizes an answer with exact timestamps and source citations. No external embedding APIs needed — everything runs locally except the final LLM call.

Why it matters: This is a complete, working implementation of retrieval-augmented generation — the same pattern used in enterprise knowledge bases, but built end-to-end from raw video to intelligent Q&A.

The problem: Running a consulting outreach operation meant juggling spreadsheets, manually sending emails, tracking call outcomes, and managing follow-ups across multiple people.

What I built: Two connected production systems. A CRM handles firm tracking, contact management, call queue with outcome logging, sustainability scoring, and CSV import/export. A separate Outreach Control Center handles automated email sending through Microsoft Graph API and SendGrid, with OAuth2 authentication, email warmup scheduling (5/day ramping to 75/day), template management, and SPF/DKIM/DMARC-compliant delivery. Both deployed on DigitalOcean with Nginx, Gunicorn, and systemd.

Why it matters: This isn't a demo — it's a multi-user production system handling real business operations daily, with proper authentication, email deliverability engineering, and production deployment.

The problem: A coaching business needed intake form responses processed, analyzed, and routed to the right team members — all manually, for every new client submission.

What I built: n8n workflow pipelines that trigger on form submission, process client data through LLM-powered analysis (OpenAI API), and route results with automated notifications. Also built 27 PowerShell scripts to automate n8n management and browser-based workflow operations.

Why it matters: Every new client submission gets processed, analyzed, and routed with zero manual steps. The same pattern fits intake forms, quote requests, and booking workflows for any small business.

The problem: Finding architecture firms that could benefit from sustainability consulting meant hours of manual Google searching, website browsing, and spreadsheet data entry.

What I built: An automated scraping pipeline using Playwright for headless browser automation, pulling from Google and Yelp. Each firm gets scored 0-10 on sustainability readiness based on their web presence, then feeds directly into the CRM for outreach. Runs on the same DigitalOcean server as the CRM with scheduled execution.

Why it matters: Shows the full data pipeline: collection, processing, scoring, and integration — the kind of automated data enrichment that scales business operations.