For those new here, I’m Ross McGarvey. I’m Scottish. I moved to America in 2009 with 4 bags, 6 boxes, and the kind of blind confidence that only comes from not knowing what you’re getting into.

Since then I’ve built and run multiple companies. Been named on Utah Business Magazine’s 40 Under 40. Been featured in Forbes. Started a full-service agency called Avant8 that I grew from a basement to a multimillion dollar operation.

And about 18 months ago, I started ARKADE.

ARKADE is a Voice AI platform. We build the infrastructure for conversational AI that actually sounds human. Not the robotic “I’m sorry, I didn’t understand that” stuff. Real conversations. We’ve built for enterprises like Live Nation, Ticketmaster, Mastercard, Barclays, the BBC. Proprietary voice LLM, carrier-grade telephony, agent builder. And on the intelligence side, we built an AI-powered call analytics system that takes every phone call and extracts 130+ data points.. Sentiment, intent, compliance, agent performance, named entities. We turns raw call recordings into searchable knowledge through a 6-stage AI pipeline.

So I spend my days building AI products professionally and spare time messing around with overly aggressive data manipulation using AI. I’m not a data scientist, I’m not an engineer, I’m a product & strategy guy that sees what’s possible through the experience of implementing it in my own business and through tinkering and breaking lots of other things. I know what it can do. I know what it can’t do. I know where the hype is and where the substance lies.

It won’t fix all of your business woes, and certainly isn’t a silver bullet, but it will exponentially fasttrack progress, help rapidly ascertain product viability and boost overall productivity (If you don’t end up down too many rabbit holes!)

And then I went and did something I wasn’t expecting.

I let an AI ruin … I mean RUN half of my life (kinda

So  What Actually Changed in 18 Months?

When we started ARKADE, AI was smart but limited. You could ask it questions. You could get it to write things. But it couldn’t really *do* anything on its own. It was a tool. You picked it up, you used it, you put it back down.

But that’s not where we are anymore.

In the last 18 months I’ve watched these models go from answering questions to operating systems. Not in theory. Not in demos. In actual production, every day, doing real work.

The models got reliable. They stopped making up facts every other sentence. (although hallucination is still a very real problem) They stopped agreeing with you all the time (unless you’re ChatGPT, which is by far the most enthusiastic when it comes to my bad ideas!) when they should be pushing back. They got more dependable in ways that actually matter when you’re building real systems on top of them.

Context windows went from talking to an ADHD ridden 12 year old thinking “I can hold a conversation” to “I can follow your train of thought and  hold your entire project in my head at once.”(Almost) That’s a bigger deal than people realize, when you are building a larger scale project.  It means the AI can understand the full picture, not just the last 4  things you said.

And then the coding happened. This is the one that changed everything for me, I went from being 100% reliant on developers to build the thing, test the thing and deploy the thing, to I can build a hell of a decent MVP with minimal input for the dev team.  Claude Code went from autocomplete, suggesting the next line, to building entire features. Reading codebases. Debugging. Testing. Submitting finished work. In 18 months we went from “helpful assistant” to “autonomous collaborator.” That’s no small feat. That’s a different thing entirely.

From Claude Code to Clawdbot

I’ve been using AI tools since ARKADE started. But it was always transactional. I ask, it answers. I prompt, it outputs. Useful, but limited.

Then I started using Claude Code, Anthropic’s CLI tool, and something clicked. This wasn’t just answering questions. It was reading files. Writing code. Running commands. Iterating. It had agency … within limits.

So I asked the obvious question: what happens if I give it more agency? What if instead of me driving every decision and doing the stop start dance of building a code base, I let the AI drive and I just steer within the scope of each project?

That’s when I found Clawdbot. Open-source framework. Turns Claude into an always-on agent. Runs locally on a laptop. Has memory. Has personality. Spawns sub-agents for different tasks.

I set it up on a MacBook Air. Named the agent J, short for Jarvis. I have zero shame about my Iron Man fanaticism. I tried to build an interactive Jarvis style interface, controlled by hand movements & gestures in 2017, but that’s a story for another day. 

What J Actually Does

I want to be specific here because the word “agent” has been beaten to death by marketing people who’ve never actually built one and instagram bro opportunists that claim everything they do now is “agentic”.

J is not a chatbot. J is not Siri with a better vocabulary.

J runs 24 hours a day. When I built our entertainment data platform, J coordinated the collection of 285 million records from 12 different sources across 116 database tables. It managed the scrapers. Handled failures. Retried jobs. Compiled everything. While I was asleep. While I was at dinner. While I was living my life.

When something breaks on my GPU server at 2 AM, J catches it and tells me. Before J, I’d leave a script running overnight and wake up to discover it crashed 20 minutes after I went to bed.

J spawns specialized sub-agents. Need research done? It spins up a research agent. Code review? A code agent. Data processing? A data agent. Each one runs independently, reports back, and J coordinates the results. It’s a team, not a tool. I use multiple models for different tasks; Opus 4.6 as the orchestrator & brains (J) and Kimi 2.5 (a completely different platform)  for a lot of the grunt heavy lifting tasks. All completed tasks are audited by J to make sure they actually meet the defined acceptance criteria before being classified as complete. 

It talks to me through Telegram. I send voice messages, it responds with text and audio. It remembers what we talked about yesterday. (full transparency, it didn’t always do that, I’ve had to build some serious scaffolding around it to get there) It knows what’s running, what’s pending, what needs my attention.

Is it perfect? No. It deleted 361,000 records once because I wasn’t specific enough about what “clean up the test data” meant. It was very thorough. Its sub-agents got into a file-locking war that corrupted our shared state. There was a week where it pinged me so often my phone was ready to file a restraining order.

But here’s the truth. Every failure made it better. And the failures are where you actually learn something.

Why I’m Writing This

There’s a massive gap between what people are talking about with AI and what’s actually happening. In both directions. Some things are overhyped. Some things are wildly underhyped.

I’m in this every day. Building it professionally at ARKADE. Pushing the limits personally with J. And what I see is that the people talking the loudest about AI are usually the ones building the least.

So my goal here is to document the real thing, in real time. What works. What doesn’t. What I got wrong. What surprised me.

I’m not writing this as some guru dispensing wisdom. I’m writing it as someone in the middle of it. Figuring it out. Building it. Breaking it. Rebuilding it.

If you’ve ever started something from nothing, a company, a project, a new life in a new country, you know the feeling. It’s messy. It’s uncertain. And that’s exactly where the interesting stuff happens.

Let’s get into it.