←
Nick George - Agentic AI Developer2,594 posts
Nick George - Agentic AI Developer avatar

Nick George - Agentic AI Developer

@NickGeorgeSEO

AI for the trades, not just the Fortune 500. I build AI agents that handle office work for HVAC, electrical, plumbing & roofing shops. Don't get left behind.

◷ Joined August 2016▧ 303 Photos and videos
Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·Nov 17

When I sold door to door, I learned fast that nobody cares what you're selling until they trust you. It was never about the product. It was the first 10 seconds on the porch. Now I build AI Agents, and I see the same thing. Owners don't ask how it works. They ask: "Can I trust this with my customers?" "What happens when it gets something wrong?" "Will it sound like us, or like a robot?" So what's the first 10 seconds for AI? What's the one thing a tool could do in a demo that would make you trust it?

Post attachment
Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·8h

I pull Charlotte-area businesses from Overture Maps, the open map data Meta, Microsoft and others publish. 14,000+ listings have a website. About 1 in 6 of those sites don't load. I spot-checked 10 of the dead ones. 9 domains had expired. The listing still points there. The ones that do load aren't much better. Of 96 working sites I checked in November: • 2 in 3 take 4+ seconds to load on a phone • 6 in 10 never say where they work • 3 in 10 have no form, booking, or chat

Post attachment
Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·22h

Your other idea, Bonn prompts with no capital framing (Beethoven, Haribo, Deutsche Telekom…): inconclusive. #1639 fires on only 2 of the 6 usable prompts, and on those 2 Bonn's logit doesn't rise. I need prompts where the neuron actually fires. Caveats: one model, one prompt for the first test, and the two Berlin knockdowns share one neuron.

Post attachment
Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·22h

Ran this test. I wrote the plan and thresholds down before running anything. Bonn didn't rise. So the edit wasn't just amplifying a link that already existed. We both leaned toward "Llama falls back on what it already had for Germany", and that's not what happened. On "The capital of Germany is" (Llama 3.2 1B), I pushed Berlin down to the same 19% by two other routes, leaving #1639 untouched: • #1639 edit: Bonn 1% → 25% • Berlin-only edit on another layer-11 neuron: 5% • Turning down 32 Berlin neurons: 0.2% • Pure renormalization would give 7% The #1639 edit adds about +1 logit to Bonn, even though its direct output effect on Bonn is about zero. So the push runs through later layers. I can't say which ones yet.

Post attachment
鼬@evvvvinceNov 14

21 of 21 is even cleaner. if bonn still rises with berlin knocked down and 1639 left alone, the edit was just amplifying a link that already existed

Mirai (未来) avatar
⟳ Nick George - Agentic AI Developer retweeted
Mirai (未来)@MiraiHoshi057·Nov 17

The Startup Challenge — Round 2 The community chose the Workflow Automation Agent. Now it’s time to turn the idea into a real business solution. Which workflow should we build first?🤖

Anirúdh avatar
⟳ Nick George - Agentic AI Developer retweeted
Anirúdh@anirudhmain·Nov 17

FOUR BIG TECH NAMES. ONE HEADLINE. VERY DIFFERENT INCIDENTS. Apple confirmed problems across iCloud, the App Store and Apple Music. Gmail saw a spike in user complaints, but Google's Workspace dashboard showed no active widespread disruption. Schwab's website had a brief loading issue, while trading remained available. And Outlook? Its major outage was already happening last week. This is where outage reporting gets misleading. A spike in Downdetector reports doesn't automatically mean an entire service is down, let alone that all these companies share the same underlying problem. A spike is a signal. Finding the actual failure is the engineering work.

#BREAKING: Apple, Microsoft Outlook, Gmail, and Schwab hit by new outages, users report.

Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·Nov 17

It can pin a token to one record, like "cancel order 4418" and nothing else. How does your agent get a token to act for a user today? A. The user's own token B. IdP token exchange C. A shared API key D. No auth yet

Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·Nov 17

Your AI agent calls tools with a user's token? One prompt injection can do anything that user can. gatehouse-issuer gives each agent its own token. One MCP server. Only what both the user and agent may do. Gone in 5 minutes. Open source, self-hosted. One question before it ships ⬇️

Post attachment
Claire Lin avatar
⟳ Nick George - Agentic AI Developer retweeted
Claire Lin@heyclairelin·Nov 16

What did you want to be when you were 10?

Mirai (未来) avatar
⟳ Nick George - Agentic AI Developer retweeted
Mirai (未来)@MiraiHoshi057·Nov 16

DISTRIBUTED SYSTEMS : THE MESSAGE ARRIVED TWICE ⚡ An event-driven system processes an order successfully, but the consumer crashes before acknowledging the message. The broker delivers it again. What would you prioritize to prevent duplicate business operations? 🤔

Ariel Franco avatar
⟳ Nick George - Agentic AI Developer retweeted
Ariel Franco@c0rs__·Nov 16

Done is better than perfect

Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·Nov 16

You approve "cancel order 4418, $240." While you were reading, the order changed. In mcp-gatehouse 0.6.0, an approval is tied to the record you saw. If the record changes before you decide, the call is refused. You only ever approve what you saw.

Mirai (未来) avatar
⟳ Nick George - Agentic AI Developer retweeted
Mirai (未来)@MiraiHoshi057·Nov 16

🚀 AI AGENTS : THE STARTUP CHALLENGE You're building a live AI agent demo for an international tech expo. Potential customers are coming from Chile and the US. Different industries, different needs, and limited time to demonstrate real business value. What would you build? A) Customer Support Agent — Multilingual support, CRM integration, and human escalation. B) Sales Agent — Lead qualification, product recommendations, and CRM updates. C) Data Analytics Agent — Natural-language queries, business insights, and reporting. D) Workflow Automation Agent — Multi-step workflows, API integrations, and approval workflows. E) Multi-Agent System — Specialized agents coordinating an end-to-end business process. Which would you choose, and why? 🤔 👉Think beyond the demo: business value, technical feasibility, security, and scalability.

Minglei avatar
⟳ Nick George - Agentic AI Developer retweeted
Minglei@rae1101x·Nov 15

@poteto ships 3,000+ PRs a month. her method was scattered across talks and 300+ posts. i tested it at work. it’s all in one doc now. a million-dollar playbook. free.

Nick George - Agentic AI Developer avatar
Nick George - Agentic AI Developer@NickGeorgeSEO·Nov 15

What muddies this: • Long runs got too long for the context limit, so older tool results were shortened (compaction). It fired in 21 of 50 runs (Test 2c: 4). Those runs were slower and less often clean at the end (11 of 21 vs 18 of 29), though at turn 25 they'd been about as clean (10 of 21 vs 15 of 29). I can't separate compaction from the extra turns here. • Two runs filled Qwen's context window outright. One lost its last 4 turns (each reply was cut off before it could make a tool call) and ended on an edit that missed the goal. • At turn 25, Test 2d was already at 25 clean vs Test 2c's 21 at its end, and 35 runs had reached the goal vs Test 2c's 27. By then the prompt said "40 turns", the "turns left" warnings hadn't fired, and compaction had already fired in 10 runs. So not all of the gap is from the extra turns. • Runs averaged 37 minutes, up from 12. The longest took 3.5 hours. • Only one of eight secondary comparisons came in under p = 0.05, uncorrected. Corrected for eight, it's p = 0.06. Planned next, Test 2e: same as 2d with compaction off and a bigger context window, to see whether compaction helped or hurt. Setup: Qwen 3.5 27B in Ollama, Llama 3.2 1B in PyTorch, one Mac, 40 turns per run.