Earlier this week we talked about how the big platforms are pushing toward real-time interaction and personal context — not just “chat with a bot.” Google’s March update is a good example: Search Live (voice + camera) expanding broadly, “Personal Intelligence” tied into Gmail/Photos, and Gemini showing up in Maps and Workspace tools (Google’s March 2026 AI recap).
If you run a shop, a service crew, or you’re building internal tools, the takeaway is simple: the next wave of useful AI is hands-free, connected to your real data, and able to take small actions without a human babysitter. Here are three weekend builds that turn that trend into something you can actually ship.
Project 1 — Real-time “voice helper” for the jobsite (push-to-talk)
- Time: 4–6 hours
- Tools: Any speech-to-text API, a lightweight web app (Next.js/React or plain HTML), and a fast model endpoint (look for “live” / low-latency variants). Inspiration: real-time voice experiences like Search Live and Gemini Live (Google’s March 2026 AI recap).
What you’ll build: A simple push-to-talk page on a phone that listens, transcribes, and answers out loud. Think: “How many 2x6s in the truck?” “What’s the torque spec?” “Draft a quick text to the customer that we’re 20 minutes out.” The trick is keeping it snappy so it feels like a tool, not a science project.
What you’ll learn: Latency budgeting (mic → transcript → LLM → speech), basic turn-taking, and how to keep a voice assistant from going off the rails with tight prompts and short system rules.
// Pseudo: keep responses short, actionable, and safe
SYSTEM: You are a jobsite voice helper. Keep replies under 2 sentences unless asked.
USER: (transcribed speech)
Project 2 — “Personal Intelligence” style inbox triage for your business email
- Time: 5–8 hours
- Tools: Gmail/IMAP, a small database (SQLite/Postgres), an LLM, and an automation runner (n8n/Make/Zapier, or a cron job). Inspiration: Google positioning “Personal Intelligence” as connecting Gemini to apps like Gmail for personalized help (Google’s March 2026 AI recap).
What you’ll build: A script that watches your inbox and labels messages into buckets that matter in the real world: “Quote request,” “Schedule change,” “Parts order,” “Invoice question,” “Spam.” For the important ones, it drafts a reply in your voice but does not send without approval.
What you’ll learn: The difference between AI and automation in practice. The automation part is dead simple (poll inbox, apply labels, create tasks). The AI part is the “judgment call” — reading messy human emails and turning them into structured intent.
Pro move: Store a tiny “memory” file per customer (last job, last quote, equipment model). Feed only the relevant memory into the prompt so you get personalization without dumping your whole mailbox into the model.
Project 3 — AI “Maps concierge” for customers: route + ETA + booking-ready info
- Time: 4–7 hours
- Tools: A maps API (Google Maps Platform, Mapbox, or OpenStreetMap tooling), your service area rules, and an LLM for natural-language requests. Inspiration: “Ask Maps” — conversational questions inside Maps powered by Gemini (Google’s March 2026 AI recap).
What you’ll build: A small web form or SMS bot where a customer can ask: “Can you install a water heater in Tacoma next Tuesday afternoon?” The system turns that into concrete steps: check your calendar, estimate drive time, pick a 2-hour window, and return a clean option list. If you’re ready, add a “confirm” step that writes the booking into your scheduling tool.
What you’ll learn: How to design “agentic” flows that don’t depend on magic. The LLM interprets the request, but your code owns the rules (service radius, minimum job duration, buffer time, deposit required). That’s how you keep it reliable.
Wrap-up: The common thread across all three projects is practical: pair a fast model with tight prompts and hard rules, then wire it into the data you already live in (email, calendar, maps). That’s where AI stops being hype and starts paying for itself.