The hype cycle around AI agents has reached a deafening pitch. Viral demos claim agents are replacing entire software departments. But behind the noise, there’s a quieter, more durable reality: the AI-native solo operator.
I’m not a software developer. My background is in Solutions Consulting for the energy sector. Yet, as of May 2026, I manage three revenue streams, a dozen automated intelligence pipelines, and a content engine producing high-signal market analysis—all powered by a stack of agents I direct, not code.
This post isn’t about what AI might do. It’s an honest inventory of what I’ve actually built, what’s running on my desk right now, and what failed miserably this month.
The Hardware Context: The Desk-Side Data Centre
Everything starts with the hardware. Many operators make the mistake of running their entire stack on cloud APIs. While I use those for complex reasoning, the "gravity" of my operation is local.
[Section removed — unverified claims. To be expanded with verified data.]
Why local inference matters:
- Cost: My automated jobs run thousands of times a month. At cloud API prices, a simple grid-monitoring agent would reportedly cost significant amounts per month [citation needed]. Locally, it costs the price of electricity.
- Privacy and agency: My Obsidian vault—my "second brain"—contains a decade of professional context. Syncing that corpus to a third-party cloud every time I want a summary isn’t an option.
- The always-on buffer: When the internet goes down or a major API provider throttles traffic, my local agents keep collecting data and synthesising notes.
The Agent Stack: My Digital Workforce
I don’t have a "chatbot." I have an agent stack, and I distinguish between agents based on their cognitive profile and workflow role.
Hermes Agent (The Executor)
[Section removed — unverified claims. To be expanded with verified data.]
Claude Code (The Engineer)
When I need to update the database schema for my energy comparison site or debug a Next.js component, I hand the keys to Claude Code. It excels at multi-file repository work that would overwhelm a general-purpose agent.
The Telegram Gateway
I’m rarely at my desk. My stack is accessible via a custom Telegram bot. If I’m in a meeting and spot a market shift, I can message Hermes to queue the work. By the time I’m home, the output is in the vault.
Five Exec Team Personas in Open WebUI
CEO, Analytics Chief, Marketing Chief, Product Manager, Tech Lead—each persona is loaded with specific context and briefed to challenge decisions from their lane. This isn’t roleplay; it’s structured dissent.
The Automation Engine: 15 Silent Workers
The most important part of my operation isn’t what I do; it’s what happens while I’m at my day job. I currently maintain 15 Python-based launchd jobs—scheduled tasks running on the Mac Mini around the clock.
Intelligence Collection
Every few hours, agents reportedly snapshot the GB Electricity Grid (BMRS data), scrape Polish Energy Regulatory Office tariff updates, and pull from 27 RSS sources covering energy, digital finance, and AI [citation needed]. All data lands as structured Markdown in Obsidian.
Knowledge Synthesis
At 02:00 every morning, a knowledge synthesiser script processes everything collected in the last 24 hours. It maintains a living "wiki" of nine topics—energy markets, stablecoins, AI agent architecture, and more. It also generates content briefs for three streams: this blog, the Polish energy site, and the UK homeowner site in development.
The Content Flywheel
Raw ideas in a "Seeds" folder are expanded into developed topics, then drafted posts, then a review queue. My role is the final filter before publishing—I review, not generate.
The Three Revenue Streams
Diversification is the strategy. By using AI to lower startup costs, I can run multiple bets simultaneously without the overhead of a team.
Lumen Pro (pickingsolutions.tech)
The flagship. A market intelligence platform at the intersection of energy, digital assets, and financial infrastructure. With reportedly over 140 posts, it serves as a portfolio and lead-generation tool for consulting work [citation needed]. It’s not currently a product push—real user feedback ("I don’t know what this site is for") led to a strategic pivot. The focus now is on building what works before selling what doesn’t.
ProstEnergia (prostenergia.pl)
A live Polish household energy price comparison site. A "Go-Compare" model for a market transitioning to dynamic pricing. The database reportedly covers 60+ suppliers and 700+ tariffs, managed by agents that scrape and normalise data daily [citation needed]. This stream is generating real interest from affiliate partners.
HomeEnergyForge (in design)
The newest project. A standalone site for UK homeowners, repurposing energy tools built for Lumen Pro into a consumer-friendly package. Domain acquisition and design brief are next.
The Monthly Failure Log
Innovation is controlled failure. This month, two lessons stood out.
The Batch Drafting Problem
I tried running 10 content drafts in a single agent session. The results were suspicious—fabricated figures, false corporate announcements, and acquisition prices that didn’t exist. The issue? As the context window fills, the model "hallucinates" to fill gaps. It’s essentially confabulation.
The fix: one post per agent session, with explicit instructions: "Do not invent figures, prices, or corporate announcements. If you cannot verify a number, use directional language or omit it." Quality improved immediately. The lesson: batch context is an anti-pattern for grounded accuracy.
The Hardware Ceiling
[Section removed — unverified claims. To be expanded with verified data.]
The Compounding Logic: Why This Matters
The goal isn’t to "use AI." The goal is to build a system where manual overhead per unit of output decreases each month.
In a traditional business, doubling content output or tracking a new market means hiring or working more hours. In an AI-native operation, it means writing a new script or creating a new agent skill. Each automation is a permanent reduction in future workload.
As of May 2026, my manual work is restricted to two functions: strategy (deciding what to build and prioritise) and review (ensuring the agents haven’t lost the plot). Everything else runs on schedule.
That’s the state of the build. It’s iterative, occasionally broken, and entirely manageable by one person. The transition from "full-stack developer" to "full-stack operator" isn’t coming—it’s already here for anyone willing to build the stack rather than just talk about it.
This is the first post in the Building in Public series. Future posts will cover specific decisions, tool evaluations, and the metrics behind each stream.
