How I Built an AI Control Centre That Fits in My Pocket
Last week I was standing in a supermarket queue when I pulled out my phone, told an AI to refactor a codebase, watched the results stream back in real time, paused the task to add a clarification, and resumed it before I reached the till. No app store download. No subscription. Built from scratch on my own machine. Here is how I did it and why it matters for any business thinking about AI.
A full AI coding assistant, accessible from any phone, anywhere
In This Article:
- 1. The Problem With Most AI Tools
- 2. What I Actually Built
- 3. The Technical Detail (For Those Who Want It)
- 4. What This Means for Your Business
- 5. The Skills Required
- 6. Where to Go From Here
The Problem With Most AI Tools
If you have spent any time looking at AI products in 2026, you will have noticed something. They fall into two camps. On one side, you have chatbots. They are easy to use, but they forget everything the moment you close the window. They cannot do anything beyond generating text. They have no memory, no ability to take action, and no way to integrate with your actual work. They are toys dressed up as tools.
On the other side, you have enterprise AI platforms. These are powerful, but they come wrapped in six-figure contracts, months of implementation, and a sales team that wants to take you to lunch before they will even show you a demo. For a small or medium business, these are out of reach. For a solo developer or consultant, they are absurd.
The gap between these two extremes is enormous. And it is exactly where the most interesting work is happening. What if you could have the power of an enterprise AI system, running on your own hardware, controlled from your phone, and built to do exactly what your business needs? That is what I set out to build.
What I Actually Built
The system I built is a bridge between Telegram, the messaging app, and a full AI coding and reasoning engine running on my own machine. It is not a wrapper around ChatGPT. It connects to Claude Code, which is an AI that can read files, write code, run terminal commands, search the web, and reason through complex multi-step problems.
From my phone, I can send a message like "review the authentication module and fix any security issues" and watch the AI work through the problem in real time. As it reads files, runs tests, and writes code, each step streams back to my phone as it happens. I am not waiting for one massive response. I am watching it think.
Key Capabilities:
- ✓ Real-time streaming: Responses arrive as they are generated, not as one block at the end
- ✓ Session continuity: The AI remembers our entire conversation history across messages and restarts
- ✓ Stop and redirect: I can halt the AI mid-task and give it new instructions instantly
- ✓ Pause and update: I can pause a running task, provide additional context, and let it resume with that new information
- ✓ Command menu: A full set of commands for common tasks like code review, committing changes, and running test suites
- ✓ Image analysis: I can send photos and screenshots for the AI to analyse
- ✓ Reply threading: Replying to a specific message includes that context automatically
- ✓ Authentication: Locked to my Telegram account only. Nobody else can access it.
The entire thing runs on a Mac Mini sitting on my desk. There is no cloud service I am renting. There is no vendor who can change the pricing or shut it down. It is mine.
The Technical Detail (For Those Who Want It)
If you are not technical, feel free to skip to the next section. But if you are curious about what sits underneath, here is a brief tour.
The architecture has three layers. The first is a Telegram bot built with the Telegraf framework in Node.js. This handles all the messaging, authentication, and the command menu you see when you type a forward slash. The second layer is a session manager that maintains conversation state. Every message you send continues the same session using a resume mechanism, so the AI has full context of everything discussed previously. The third layer is the AI engine itself, invoked as a command-line process with streaming JSON output.
One of the more interesting design decisions was the command registry. Instead of scattering command definitions across multiple files and handlers, every command is defined in a single array. Each entry specifies the command name, its description, and how it should be handled. That same array is used to register commands with the Telegram menu, route incoming messages, and generate the help text. One source of truth, three uses. Adding a new command is a single line of code.
The stop and wait system was particularly satisfying to build. When you type "stop", the running AI process is killed immediately. When you type "wait", the system enters a paused state where your next message is captured as additional context, and the original task resumes with that new information folded in. This turns a one-directional interaction into a genuine conversation, even when the AI is mid-task.
Streaming was essential. Most AI interfaces make you wait for the complete response before showing anything. For complex tasks that might take thirty seconds or more, that is a terrible experience. Instead, this system parses the streaming JSON output and sends each text block and tool invocation to Telegram as it happens. You see the AI working in real time.
What This Means for Your Business
Here is the part that matters most, regardless of whether you care about the technology underneath.
Most businesses that are "adopting AI" in 2026 are doing one of two things. They are either giving their staff access to ChatGPT and hoping for the best, or they are paying a large consultancy to bolt on an AI layer that nobody really understands. Both approaches have the same problem: the AI does not know anything about your business. It does not know your processes, your data, your customers, or your constraints.
A custom AI integration is different. It is built around the way you actually work. It has access to the systems and data that matter. It can take real actions, not just generate suggestions. And because it is yours, it improves over time as it learns more about your context.
Real Examples of Custom AI Integration
A therapy practice
An AI assistant that handles scheduling queries, sends appointment reminders in the right tone, drafts initial assessment summaries from intake forms, and flags safeguarding concerns in client notes. All running privately, all GDPR compliant, all tailored to how that specific practice works.
A restaurant group
Real-time competitor monitoring that tracks menu changes, pricing, and review sentiment across rival restaurants. Automated responses to reviews that match the brand voice. Daily briefings delivered to the owner by text message every morning.
A trades business
Automated quoting based on photos of the job site. The customer sends a picture, the AI analyses it, cross-references the materials database, and generates a quote. The business owner approves or adjusts it from their phone in thirty seconds.
A recruitment agency
An AI that reads incoming CVs, matches them against active roles, scores them based on criteria the agency defines, and sends a shortlist to the recruiter every morning. No more spending hours sifting through applications.
None of these examples require enterprise budgets. They require someone who understands both the technology and the business problem, and who can connect the two. That is integration work. That is what I do.
The Skills Required
Building something like this is not about knowing one tool or one programming language. It requires a particular combination of skills that is still relatively rare.
- API design and system architecture. You need to understand how to connect systems that were never designed to talk to each other. That means working with APIs, message queues, webhooks, authentication protocols, and data formats.
- Prompt engineering. The way you instruct an AI matters enormously. A well-crafted system prompt can be the difference between an AI that is useful and one that is useless. This is not about tricks or hacks. It is about understanding how these models reason and giving them the right context.
- State management. AI interactions are not stateless. You need to maintain conversation history, handle interruptions, manage concurrent requests, and deal with failures gracefully. This is traditional software engineering applied to a new problem.
- Fast iteration. The AI landscape moves quickly. The ability to prototype something, test it with real usage, and refine it in days rather than months is essential. This is where experience matters. Knowing what to build, what to skip, and where the pitfalls are saves enormous amounts of time.
- Business understanding. The hardest part is not the code. It is understanding what the business actually needs and translating that into something an AI can do well. This requires listening carefully, asking the right questions, and knowing when AI is the right tool and when it is not.
I have been building software professionally for over fifteen years, across IT architecture, web applications, and more recently AI systems. I also have a clinical background in cognitive behavioural therapy, which turns out to be surprisingly useful when designing AI interactions. Understanding how people think, how they process information, and how they respond to different communication styles makes the systems I build more natural and more effective.
Where to Go From Here
If you have read this far, you are probably thinking about what a custom AI system could look like for your business. Maybe you have a process that eats up hours every week. Maybe you are drowning in data that nobody has time to analyse. Maybe you have already tried off-the-shelf AI tools and found them underwhelming.
The mobile control centre I built for myself is a proof of concept for a much larger idea: that AI should be built around your workflow, not the other way around. The businesses that will get the most value from AI in the next few years are not the ones that adopt the most popular tools. They are the ones that build systems tailored to how they actually operate.
I work with businesses to figure out where AI can genuinely help, then I build it. No six-month discovery phase. No slide decks. Just a conversation about what you need, followed by working software.
Let's Talk About AI for Your Business
I offer a free initial consultation to understand your business and identify where custom AI integration could save you time, money, or both. No sales pitch. Just an honest conversation about what is possible and what is worth pursuing.
Elliot Rose
Senior developer and AI integration consultant based in Surrey, UK. Fifteen years in software architecture, clinical background in CBT, and a deep interest in building AI systems that solve real business problems.