I've been running cold email campaigns for clients for 3 years and the biggest shift I've seen isn't the tools. It's what actually gets a reply. Personalization used to mean scraping a name and company from LinkedIn, dropping it in the first line, and hitting send. "Hey {FirstName}, I noticed {Company} and thought..." That worked in 2022. It's dead now. Everyone's doing it and prospects can spot a mail merge from the subject line. What changed for me was treating personalization like actual research instead of a data field. Here's what I started doing: → I scrape the prospect's entire website. Not just the homepage. Blog posts, service pages, case studies, about page, even their contact form if it's there. → Then I feed all of that into OpenAI and have it analyze what they actually do, who they serve, and what problems they're likely dealing with. The AI doesn't just summarize. It finds the specific details nobody mentions in generic outreach. So instead of "I saw you work in logistics," the email opens with "Noticed you handle cross border freight into Mexico. Your blog mentioned customs delays eating 15% of delivery windows." That's the kind of line that gets opened because it doesn't sound like 500 other emails they got that week. The reply rates went from 2-3% with generic personalization to 8-10% with actual research. One prospect replied last week: "Your email won because you actually read our site. Everyone else sent the same template." The system I built does this automatically. Scrapes the website. Analyzes every page. Generates icebreakers that reference non-obvious details. It writes openers like a human who spent 20 minutes studying their business, except it does it for 1,000 prospects in an hour. Here's what I learned building this: Small prompt details make a massive difference. Having OpenAI shorten company names naturally (say "Stripe" not "Stripe Inc.") and reference specific pages beyond the homepage makes it feel real. The difference between "I saw your website" and "I saw your freight tracking dashboard lets customers get ETAs without calling" is everything.

Every automation you build has a round trip baked in. Claude API, OpenAI call, n8n webhook, vector DB lookup. Data leaves the device, travels to a data center, gets processed, comes back. On a good day a cached Claude call runs 600ms to 1.2s. On a bad day you are watching a spinner. Where that round trip breaks: - Real-time perception in vehicles or robotics - Industrial control loops that cannot wait on a network - Anywhere connectivity is spotty or intermittent - High-volume sensor data where shipping everything to the cloud burns bandwidth and budget Edge changes the math. Compute runs where the data is generated. Local work stays local. Cloud only sees what needs cross-site context. The split that is emerging: - Edge: real-time inference, event detection, filtering, local control - Cloud: model training, cross-site analytics, long-term storage, heavy compute On-device SLMs are usable now. Llama 3.1 8B on an M-series Mac via Ollama hits sub-100ms first token. Haiku-class reasoning is running on phones with NPUs. If your automation touches physical systems, live audio, live video, or anything latency-sensitive, you have a real placement decision to make: edge, cloud, or hybrid. The pattern I keep reaching for: route by cost of latency. If a one-second delay breaks the experience, run it local. If the step needs cross-context memory or a frontier model, send it up. Cheap decisions at the edge, expensive ones in the cloud, with the edge doing the filter so the cloud only sees signal. The architecture question has changed. Not "which cloud model do I call," but "where does each step of this pipeline need to run." What is running in your stack right now that should not be making a cloud round trip? What is stopping you from moving it?