In this episode, we interviewed Dheeraj Sharma - Cloud FinOps practice lead by day and the builder behind GenAI Unplugged, an AI newsletter, product studio and academy that teaches non-technical founders and creators to build real AI systems, not just use AI
Dheeraj has spent 20 years shipping enterprise software that has to survive governance, security and real budgets - and he runs a one-person business on a six-layer AI operating system, with 12 live products and a Substack Bestseller badge earned in his first six months
His core argument: the demo is the easy 20% - the unglamorous 80% of cost, monitoring, error handling and security is what separates a tool that impresses in a meeting from a system a business can actually depend on
Key takeaways:
Day 47 is the real test - a demo runs beautifully in the boardroom, but whether it survives cost, a bad input at 2am and unattended production is the 80% everyone skips
AI collapses the build, so the bottleneck moves to judgment - when code takes days instead of weeks, the constraint becomes review and knowing what not to ship
AI amplifies judgment, it doesn't replace it - it's spectacular at implementation and useless at knowing what's worth implementing for your domain, and that part is still yours
Governing agents becomes the budget line - within 12 months the premium shifts from building agents to operating them: cost, evaluation, monitoring, failure handling
🔗 Connect with Dheeraj
Who are you and what do you do?
I lead a Cloud FinOps practice at an IT services firm, where I have spent 20 years shipping enterprise software that has to survive governance, security, and real budgets. Nights and weekends I run GenAI Unplugged, where I build AI products and systems in public and teach non-technical founders and creators to build real systems, not just use AI. I live on both sides of the AI conversation: the enterprise world that asks "will this survive production?" and the solo world that asks "can I ship it this weekend?"
What problem did you see that everyone else was missing?
Everyone was shipping AI demos. Almost nobody was owning what happens on day 47, when that demo is running unattended in production. I learned this the expensive way: an automation I built kept silently retrying on errors and quietly ran up over $300 in unexpected API bills before I noticed. The demo is the easy 20%. The unglamorous 80%, cost tracking, error handling, monitoring, security, is what separates a tool that impresses in a meeting from a system a business can actually depend on. I built an entire series, Demo to Dependable, around that 80%, because it is exactly where real money is won and lost. Enterprises have known this about software for decades. With AI, everyone suddenly forgot.
AI is spectacular at implementation and useless at knowing what is worth implementing for you or your domain or context, and that part is still yours.
Walk us through one concrete way your work changes what companies actually ship
When I built SubflowAI, a Chrome extension that touches people's live publishing accounts, the code was the fast part. Claude Code wrote most of it in days. What decided whether I could actually ship was the review. Before release I had Claude spawn parallel review agents, each auditing one thing: one on permissions and content-script scope, one on every network call, one on injection and API-key exposure, one on performance under 10,000-plus scheduled items. I did not ship a single line until that review came back clean. That is the real shift AI creates: it collapses the build from weeks to days, so the bottleneck moves entirely to judgment, review, and knowing what not to ship. The companies getting burned are the ones who sped up the building but not the reviewing. They ship faster and break more.
What's the most common thing senior leaders get wrong about AI?
They treat AI as a replacement for judgment instead of an amplifier of it. I watch leaders ask AI "what should we build or automate?" and expect a decision. AI does not give you a decision, it gives you every option, confidently. Ask it to add a settings page and it hands you thirty toggles you should never ship. AI is spectacular at implementation and useless at knowing what is worth implementing for you or your domain or context, and that part is still yours. The second mistake is buying the demo. A polished demo in a boardroom is 20% of the job. Whether it survives cost, security, and a bad input at 2am is the other 80%, and no demo shows you that. The leaders getting real ROI pair AI's speed with human taste. They do not outsource the taste.
It collapses the build from weeks to days, so the bottleneck moves entirely to judgment, review, and knowing what not to ship.
What's in your AI stack? The one tool you rely on every week?
My whole stack is one architecture: six layers with a single orchestrator in the middle, built so I am never hostage to one vendor. The orchestrator is Claude Code, with Codex sitting behind it as a backup brain, same files, same commands, different vendor, so an outage never stops me shipping. Around it: Gemini and Tavily for research on a free-first, paid-last rule. n8n on Oracle Cloud's free tier as the plumbing. Notion and a local SQLite database as the memory. A fleet of Claude skills and specialist sub-agents on top. The one I lean on every single week is Claude Code as the orchestrator: one plain-English command reads my business context, researches, drafts, generates the image, and drops the finished asset in the right folder. Six layers of context, one command, one output. That is how one person ships twelve products and three articles a week.
What does your work actually look like day to day?
The headline is "Solo AI Systems builder." The real version is margins. I have a full-time day job leading a Cloud FinOps practice, so GenAI Unplugged happens at 5am, in lunch breaks, and after my family is asleep or before they wake up. Three to four hours a day, almost never in one block. It only works because I run three layers: automate, batch, compress, and a contentOS I built carries what I used to hold in my head. The system tells me what is next, so I just follow. I batch instead of context-switching, product features in one sprint, articles in another. And I use AI to compress small windows, a single evening now ships what used to take a week. It is not glamorous. It is early mornings and a system that does the remembering for me. I never chase new systems or tools unless there is any significant implication on what is already working for me.
Where is your field in 12 months - one specific prediction?
Within 12 months, "we built an AI agent" will be worth nothing, and the entire premium shifts to operating them. My specific bet: the fastest-growing AI budgets next year will not be for building agents, they will be for governing them, cost, evaluation, monitoring, failure handling. The FinOps discipline, applied to AI. The companies that win in 2027 will not be the ones with the most agents, they will be the ones who can prove their agents are cheap, safe, and reliable enough to leave running. I would put money on "agent governance" or "AgentOps" becoming a named line item in enterprise budgets by mid-2027, the same way cloud cost management became one 4-5 years ago just after COVID. I have watched this exact movie before, in cloud. The sequel is already in production.
Where should readers find you, and what’s the first thing they should join or read?
Start at genaiunplugged.substack.com, that is the front door to everything I publish, and my YouTube channel if you prefer to watch. If you engage with one thing, make it my Demo to Dependable work. It is the most useful for anyone actually putting AI into production rather than just experimenting. It walks through the parts everyone skips: what your automation costs when it misbehaves, how to catch failures before your users do, how to make a weekend build survive a real workload. If you lead a team and you are tired of impressive demos that never quite become dependable systems, that work was written for you. I write and film in plain English, with real numbers from things I actually shipped, and occasionally broke. No theory I have not run myself.
AI Central Voices is where the AI Central team sits down with the founders, executives, and builders shaping AI - going behind the scenes of how they operate, what they're betting on, and where the industry goes next.
Want to be featured, or have an event you'd like us to cover? Reach out at [email protected]
