What I built, what broke, and what I learned. Building with AI since May 2026, about 100 hours so far.
My first client website: The Garrett Company
Built a full website for a third-generation real estate brokerage in Franklin, TN, founded in 1947. Live preview.
Researched the company's real history, team, 21 recent home sales and 8 commercial deals, and put all of it in one data file so nothing on the site is made up.
Hero of real Nashville drone footage (a vertical clip on phones), a self-drawing legacy timeline, and a filterable sales gallery.
Wrote a typo-tolerant location search over every Middle Tennessee city and ZIP (Damerau-Levenshtein matching), so "frsanklin" still finds Franklin, TN.
Lesson: real details beat placeholders. The company's own story (same office since 1966, "small on purpose") was better copy than anything I could invent.
What broke: my first upload to Netlify sent only index.html, so the site showed up as plain text. Uploading a zip of the whole folder fixed it.
A waypoint graph, A*, steering with a yield rule, spot reservation and a hard 0.55 m no-overlap rule took clipping from constant to zero.
Week of Oct 12: Forge Labs, a 3D lab, and going public
Set up Forge Labs: a team of AI research agents (historian, psychologist, quant, physicist, venture architect and more) run by a director agent, all writing into one Obsidian vault that syncs to GitHub automatically.
The agents wrote 600+ evidence-rated notes: 100 psychology principles (with honest replication ratings), 200 math and quant principles (every worked example checked in Python) and 150 physics principles so far.
Built a 3D Science Lab in Three.js where the agents show up as scientists in 9 themed rooms, with A* pathfinding through doorways, a walk cycle, sitting and typing, conversations, and collision-free steering.
Released Vault Atom (an Obsidian plugin, MIT) and Forge AI (a local AI with RAG over my notes, Apache-2.0).
What broke: running four agents plus a local 8 GB model at once made my 16 GB Mac glitch. Now agents run one at a time.
Lesson: model choice matters. Switching the note-writing agents from the biggest model to a mid-size one, and checking sources with one script instead of opening every page, made each note about 10–15× cheaper with the same quality.
Lesson: a lot of famous psychology doesn't replicate (ego depletion, power posing, oxytocin and trust). Rating the evidence on everything was worth it.
September: my agency website
Designed and built a studio website with a scroll-driven animated layout (GSAP and Lenis), a pricing section and live project cards.
June: data pipeline
Connected a B2B data API to my lead agent, filtered to US businesses, and checked each website automatically. Ended up with 8,000+ qualified leads, every one with an email.
Lesson: data quality beats data quantity. Verifying emails before sending protects your sending domains.
May: my first AI agent and a whole outreach stack
Built my first agent: it finds small businesses with broken or outdated websites, scrapes contact emails and writes a personalized pitch.
Built the tools around it: a 100-point lead scorer, an AI website-mockup generator, proposal and contract generators, a reply classifier with a one-click approval dashboard, and a booking webhook.
Set up cold-email infrastructure from scratch: 15 domains with SPF, DKIM and DMARC, 42 inboxes, and warm-up.
About 44 hours in 10 days. My busiest day was 226 prompts.
Lesson: never hardcode API keys. I learned this the hard way and now use a pre-commit guard that blocks them.