2026-03-19
By Vadym · Generated with Boba, curated by me
Less than a week ago I started an experiment. What if I took a frontier AI model and wired it into everything — my email, my calendar, my codebase, my research workflow, even my voice notes while I'm at the gym?
What surprised me is how fast it came together. In under a week, what started as a simple chatbot became something I've started calling my AI operating system. Not a product, not a startup — just a personal setup that's already become genuinely useful in ways I didn't expect.
Here's what it actually looks like.
The backbone is OpenClaw — an open-source framework for building AI agents that can use tools, run scripts, and maintain context across conversations. OpenClaw gave me the foundation — the agent runtime, the tool framework, the messaging layer. But everything on top of it I built myself, mostly by talking to the AI and iterating with Claude Code. That's the part I find fascinating: the system I use daily was largely built through conversation with the AI that runs it.
On top of OpenClaw, I've built an assistant named Boba who handles a growing list of things I used to do manually. Boba runs on Claude, Anthropic's latest model. He has access to my email (via API), a shared Google Calendar, a set of automation scripts, and a voice interface through Telegram. The whole thing runs on my MacBook with scheduled jobs that fire throughout the day. And I'm still building — every week I revise something, add a new capability, or rethink a workflow. Claude Code helps me make sure everything stays stable and expandable as the system grows.
The key word is curated. Boba doesn't have the keys to the kingdom. He has access to a dedicated email address and a shared calendar — not my personal accounts. Every action that goes outward (sending an email, posting something, making a change to production code) goes through me first. Security is a feature, not an afterthought.
Voice-first workflow. I send a voice note on Telegram — from the gym, from the car, wherever. Boba transcribes it locally using Whisper, understands the request, does the work, and replies with a voice note or text. The whole loop takes about 30 seconds.
For example: I'm at the gym and I think of an improvement for a project I'm building. I record a 15-second voice note describing the change. By the time I finish my set, Boba has made the change, committed the code, and deployed it. I can see the update live on my phone. That's not a hypothetical — that's Tuesday.
Research on demand. When I want to understand something — a market trend, a competitor landscape, a technical architecture — I describe what I need and Boba creates a document. Deep research with real sources, formatted as a PDF, and sent to my email. Sometimes I queue these as overnight tasks — I describe the research before bed, and the report is waiting in my inbox by morning.
Scheduled newsletters. I have a weekly AI engineering newsletter that Boba assembles autonomously. He scans for the most important developments, writes concise summaries, and emails me the draft. I review, tweak if needed, and send. What used to take me two hours on a Sunday now takes ten minutes of review.
Calendar management. "Add a meeting Friday at noon about X" — done. Reminders set, description filled in, the right calendar selected. It sounds simple, but removing that friction from dozens of small scheduling tasks per week adds up to real time saved.
Stock watchlist. I maintain a watchlist of stocks I'm tracking. Every morning at 9 AM, Boba runs a technical analysis — price, trend, RSI, MACD, support/resistance levels — and only notifies me when something is actually actionable. No noise, no daily alerts about nothing. Silence means "nothing to do," which is usually the right answer.
Overnight task queue. This is my favorite part. I have an async job queue where I drop tasks throughout the day — research projects, code tasks, analysis work. At midnight, a runner picks up the highest-priority task and works on it autonomously. By morning, the results are in my inbox and my Telegram. The queue is persistent: if a task fails, it retries. If the runner crashes, unfinished tasks stay queued for the next night.
It's like having a junior analyst who works the night shift and never forgets what you asked for.
The compound effect. No single capability here is revolutionary. Voice transcription, email APIs, calendar integration, scheduled scripts — all of this existed before AI. What's new is that one system understands all of it in context. Boba knows what I asked yesterday, what's on my calendar tomorrow, and what research I'm waiting on. That context is the multiplier.
Voice changes the dynamic. When I started, I typed everything. Then I switched to voice notes and the usage pattern completely changed. Speaking is faster than typing, and it means I can interact with the system while doing other things — driving, cooking, exercising. The AI went from a tool I sit at to something ambient.
The overnight queue changed how I think about work. Before, if I had a research idea at 11 PM, I'd either do it right then (bad for sleep) or forget it by morning (bad for productivity). Now I describe it in 30 seconds and it's done by morning. That simple change has been more impactful than any individual AI capability.
I stopped over-engineering. Early on, I built elaborate automation pipelines. Most of them broke or weren't worth maintaining. The stuff that stuck is simple: a voice interface, a task queue, a few scheduled jobs, and a model that's smart enough to figure out the rest. The lesson is the same one from software engineering: start simple, add complexity only when it earns its keep.
I want to be honest about the limits.
Boba doesn't make decisions for me. He presents information, does analysis, and executes tasks — but the judgment calls are mine. When the AI says "the market outlook is mixed, no clear signal," the right response is to do nothing. Learning to trust that has been its own journey.
He also doesn't replace deep thinking. If I need to design an architecture or write a strategy document, AI can accelerate the research and drafting, but the hard thinking still happens in my head. The AI is a lever, not a replacement for judgment.
And the setup isn't plug-and-play. This took real engineering work to build. It's not something you install from an app store — it's a custom system built on open-source tools, tailored to my specific workflow. That's both its strength and its limitation.
I think personal AI systems like this will become normal within a few years. Not necessarily as DIY projects — more likely as products that abstract away the plumbing. But the pattern of "AI agent with access to your tools, running in the background, learning your preferences" — that's going to be table stakes.
The interesting question isn't whether AI assistants will exist. It's how much agency we'll give them. Right now my setup is heavily curated — I approve everything that goes outward. But each month, I find myself loosening the guardrails slightly as trust builds. That gradient from "tool I supervise" to "agent I delegate to" is where the real story is.
We're early. But it's already useful in ways that are hard to go back from.
By Vadym · Curated by me, built with Boba