AI Friday #2

Mike Mercuri showed off Quote Text, a chatbot he built for home service contractors to collect quote info over SMS. No dev background — just a Rev Ops guy who got tired of waiting a week for a tree trimming quote that should take five minutes. He wired together GPT, Vercel, Supabase, and Twilio to make it happen. For property lookups he uses Google Maps and Melissa Data. His development process: screenshot forms and ask GPT what to fill in. Voice commands because he types one-handed. A hashtag-based documentation system so every prompt starts with the right context. It’s in beta now.

Chris Vocke built an AI-powered response suggestion layer on top of Zendesk for a 600-ticket-a-day e-commerce and telemedicine support team. The challenge wasn’t just volume — it was language nuance, medical protocols, and getting new agents up to speed fast. His architecture stacks multiple AI layers: tagging, response generation, evaluation, and vectorization. He runs Gemini 2.5 Pro/Flash as his primary models with fallbacks, all hosted on Cloudflare Workers (with AI Gateway, D1, and a vector DB). Manager-approved responses get vectorized into a “golden responses” library. He tracks edit distance with Levenshtein Distance to measure how much agents actually change the suggestions. Accuracy: 90–100%.

Laura Williams built SPF (Sprint Product Focus) because Jira was too complicated for what she needed. It’s a personal sprint tracking tool at spf.hosette.net with side-by-side sprint views, context switching tracking, velocity analytics, and AI-powered one-on-one summaries. She built it with Lovable, which now runs Claude 4.7 under the hood. Her approach for complex features: write a spec and a milestone plan first, then build. One external user so far. She’s thinking about what to charge for the AI features.