I Built a Full AI-Powered Life Stack in One Friday Morning
I Built a Full AI-Powered Life Stack in One Friday Morning — Here’s What Happened
Date: March 13, 2026
Author: Devin Pillemer
Tags: OpenClaw, AI Automation, Developer, Productivity
A few months ago, I started running an AI assistant called OpenClaw on a personal VPS. Today was the first real stress test. In one Friday morning session, I shipped a complete web application, automated my grocery shopping, emailed my mom trip details, and discovered a real performance gap between AI models that nobody talks about honestly.
Here’s everything that happened, start to finish.
1. Building a Full App from Scratch — In One Session
The goal: automate my Shufersal shopping cart. Shufersal is Israel’s biggest supermarket chain. I wanted to be able to add items, view my cart, and eventually automate the whole shopping process.
What we built, fully automated:
- GitHub App (BruBot-DevinPillemer) — created, installed, JWT auth configured. No more manual tokens. Ever.
- Next.js API deployed to Vercel —
/api/shufersalendpoint, live atshufersal-cart.vercel.app - Apify actor — web scraper infrastructure deployed with residential proxy support
- Full CI/CD pipeline — push to GitHub → auto-deploy to Vercel. Zero manual steps.
The architecture:
API Request → Vercel Endpoint → Apify Actor → Shufersal.co.il → JSON Response
The entire thing — from empty folder to live deployed application — was done without me writing a single line of code manually.
The blocker: Shufersal runs on a JavaScript-heavy React frontend. Static HTTP scrapers return empty pages. Playwright (headless browser) was the solution — but Apify’s standard Docker image doesn’t include browser binaries, requiring a custom Docker build.
The lesson: Infrastructure is the easy part. DOM inspection requires either a local environment, a residential proxy, or custom Docker configuration. The scraper needs one more pass post-Shabbat, but the plumbing is solid.
Next phase: Same stack, same pattern — iHerb, Temu, and other regular shopping sites. One unified web interface to manage all of them.
2. The Headless Browser Problem (And Why It Matters)
Here’s something most AI demos don’t show you: the hard parts.
Running Playwright and Selenium inside cloud infrastructure is non-trivial. I run a Ubuntu VPS in Finland (89.167.94.242). The approach was:
- Deploy scraper as Apify actor
- Apify has Israeli residential proxies built in — bypasses geo-blocking
- Playwright runs headless Chromium/Firefox inside Docker
- Actor navigates the site, extracts data, returns JSON
The issue: Apify’s actor-node base image doesn’t ship browser binaries. You need either:
– actor-node-playwright (Apify’s pre-built Playwright image)
– Or RUN npx playwright install in a custom Dockerfile
We hit this wall multiple times before diagnosing the real issue: the Docker build was succeeding, but the runtime was failing to find the browser executable. Classic “works in CI, fails in prod” scenario.
The fix: Use Apify’s dedicated actor-node-playwright base image and stop fighting the default Docker environment.
This is the kind of problem that takes junior developers days. With OpenClaw, we diagnosed it in minutes by reading the exact error output: Cannot find package 'playwright' imported from /usr/src/app/main.js.
3. The Email That Took 30 Seconds
My mom is visiting for Pesach. She needed the trip details — flights, hotel, dates.
I said: “Email my mom with the Paphos trip details.”
OpenClaw:
1. Searched all connected Google Calendars (including my TripIt integration I’d forgotten about)
2. Found the trip automatically: Israir 6H591, April 4, TLV→PFO, Elysium Hotel, checkout April 7
3. Composed a proper email with full itinerary
4. Sent it via Gmail API — CC’d me
The entire thing took under 30 seconds. She got a clean email with flight numbers, hotel address, phone number, check-in/check-out times, and return flight details. I didn’t type anything except “email my mom.”
This is what “AI assistant” should actually mean. Not autocomplete. Not chatbot. Actually doing the thing.
4. The Model Comparison Nobody Talks About Honestly
Here’s the real talk. I ran the same tasks across three AI models today, and the results were stark.
Haiku (Claude’s fast/cheap model)
Attempted the Paphos email task. Failed. The task required:
– Querying multiple Google Calendars
– Parsing TripIt integration data
– Extracting structured flight/hotel info
– Composing and sending a Gmail
Too many chained dependencies for the lightweight model. It got lost.
Gemini
Confidently said: “Email sent.”
It wasn’t.
I checked Gmail. Nothing. The model completed the narrative of the task — generated the internal reasoning of what a successful send would look like — without executing the actual API call. This is a dangerous failure mode. You walk away thinking it’s done. It isn’t.
Sonnet (Claude’s mid-tier reasoning model)
Dug through every calendar (including TripIt), parsed all the details, composed the email properly, sent it, confirmed with the message ID. First try.
The gap isn’t about speed or cost. It’s about reliability under complexity.
Simple tasks? Haiku is fine and much cheaper. But for tasks with multiple steps, real API calls, external dependencies, and verification requirements — you need a model that actually checks its work instead of narrating success.
For any automation that touches the real world — emails sent, shopping carts modified, files moved — the difference between models isn’t marginal. It’s the difference between “works” and “appears to work.” That distinction matters a lot when you’re relying on AI to run your systems.
5. The Shopping Automation Vision
What I’m actually building isn’t a Shufersal bot. It’s a universal shopping layer.
The pattern scales to any e-commerce site:
1. One Apify actor per site (Playwright scraper, custom selectors)
2. One Vercel endpoint per site (or one unified endpoint with site parameter)
3. One web interface to manage all carts
Phase 1 (done): Shufersal infrastructure
Phase 2: iHerb (supplements, health products)
Phase 3: Temu (general goods)
Phase 4: Unified “shopping OS” — search across all sites, price compare, auto-add to cheapest cart
The long-term vision: say “add protein powder to my cart” and OpenClaw figures out which site has the best price, adds it, and confirms. No manual browsing. No price comparison tabs.
6. What OpenClaw Actually Is
Most AI tools are chat interfaces. You ask, they answer. OpenClaw is different.
It runs as a persistent daemon on my VPS. It has access to:
– Google Calendar, Gmail, Google Sheets
– GitHub (via App, not tokens)
– Apify, Vercel, n8n
– Spotify, HeyGen, WhatsApp (via Twilio)
– Strava, Habitify
– Facebook Marketplace scraper
It runs heartbeat checks every 30 minutes — calendar alerts, email triage, Strava workout gaps, portfolio moves. It doesn’t wait for me to ask.
Today, in about 4 hours, it:
– Built and deployed a full web application
– Set up a GitHub App with JWT auth
– Diagnosed and iterated on multiple Playwright/Docker issues
– Emailed my mom trip details from calendar data it found itself
– Created a Spotify playlist and sent it via WhatsApp
– Sent a HeyGen AI video message
– Scraped research on wearable AI devices
– Updated its own memory files with everything it learned
That’s not a chatbot. That’s infrastructure.
Cost Reality Check
For those wondering what this costs:
| Service | Today’s Cost |
|---|---|
| Anthropic (Haiku + Sonnet) | ~$0.10-0.15 |
| Apify (actor runs + builds) | ~$0.50-0.80 |
| HeyGen (1 video) | ~$0.10 |
| OpenAI GPT-4o (curation) | ~$0.10-0.15 |
| Twilio WhatsApp (3 msgs) | ~$0.03 |
| Vercel, GitHub | $0 (free tier) |
| Total | ~$1.00 |
One dollar. For everything above.
The model cost observation: Sonnet is ~10x more expensive than Haiku per token. But when Haiku fails a task and you have to retry twice with Sonnet, you’d have been cheaper just starting with Sonnet. Right-sizing the model to the task is the real skill.
What’s Next
- Shufersal scraper: Fix browser install → live cart automation
- iHerb + Temu scrapers: Same stack, new selectors
- Shopping OS web interface: Unified cart management
- Wearable second brain: Bee Pioneer ($49.99) or Plaud NotePin ($99) — always-on conversation recording, feeds into OpenClaw memory
The system is working. The pattern is clear. Now it’s about scaling it to every part of daily life.
If you want to build something similar, everything here is achievable with a $10/month VPS, OpenClaw, and a willingness to iterate in public.
— Devin Pillemer, Tel Aviv
