Frequently Asked Questions
Q: How many production agents does AIdeazz actually run on Oracle Cloud right now?
A: 11 multi-agent systems handling 4,200 daily sessions across Telegram and WhatsApp. The largest one routes 68% of traffic through Groq for speed and falls back to Claude 3.5 Sonnet only on the 9% of queries that need deeper reasoning. Monthly infra bill stays under $380.
Q: What was the single biggest executive skill that translated directly to shipping agents solo?
A: Stakeholder mapping and ruthless prioritization. As Deputy CEO I killed 60% of proposed initiatives before they reached the board. The same filter now decides which agent features get built this sprint versus parked in the backlog. Without it I would have 27 half-finished agents instead of 11 that actually work.
Q: Did your Russian regulatory and infrastructure background help or hurt the Panama move?
A: It helped with compliance and threat modeling but hurt on velocity. Russian programs required three signatures and six-month lead times. Panama lets me spin up a new Oracle shape in 11 minutes. The gap between those two realities forced me to throw away every process that assumed slow feedback loops.
Q: How long did it take before you stopped hiding the 14-year executive gap on your profile?
A: 19 months. The first 11 months I presented as a “former operator learning AI.” Conversion rate on inbound leads was 8%. Once I listed the exact pivot and the $380 monthly burn, conversion jumped to 31% and the right technical cofounders started reaching out.
Q: What is the actual cost and uptime of running multi-agent routing on Oracle with Groq fallback?
A: $0.0034 per completed session at current volume. Uptime for the routing layer has been 99.94% over the last 90 days. The only outages came from my own bad prompt updates, not from Oracle or Groq.
I shipped the first production agent 14 months after leaving the Deputy CEO role. The failure that almost killed the pivot happened in month four: I spent $11,400 and 187 hours building a “comprehensive platform” with 14 microservices before realizing zero users needed it. The lesson was expensive and immediate. Everything after that decision was shaped by constraints that most ex-executives refuse to admit exist.
What Executive Experience Actually Transfers
Budget control under uncertainty transfers perfectly. In the Russian digital infrastructure program I managed a $47M annual spend with quarterly reviews that could cut 30% overnight. The same muscle now keeps AIdeazz at $380 monthly on Oracle Cloud while handling 4,200 sessions per day. I know exactly which line items can be cut without killing the product.
Negotiation with vendors transfers. Oracle sales reps respond to the same pressure I used on Russian state contractors. When I told them I would move the entire workload to Hetzner if they could not match the shape pricing, they adjusted within 48 hours. The language is different but the power dynamic is identical.
Hiring and firing decisions transfer. I have fired three contractors in the last year. Each time the pattern was the same: they optimized for impressive demos instead of reliable uptime. The executive version of this was firing department heads who missed KPIs. The technical version is firing an agent that hallucinates pricing data in production.
The skill that surprised me most was crisis communication. When one of the WhatsApp agents started returning Russian-language error messages to Spanish-speaking users in Panama, I had 40 minutes to draft the apology, push the rollback, and update the routing table. The same template I used during the 2019 infrastructure outage in Moscow worked with almost zero modification.
What Was Completely Useless
PowerPoint governance theater. I used to spend 18 hours per quarter preparing 47-slide decks for the board. That skill has negative value when you are the only person who needs to understand the decision. I now track everything in a 14-row Notion table. Anything longer is waste.
Committee-based risk management. In government-adjacent programs every decision required sign-off from legal, security, and three directorates. The result was six-month cycles and diluted accountability. When I tried to replicate even a light version for AIdeazz, the first agent took 43 days to reach production. I killed the process the same week.
Corporate branding discipline. The obsession with consistent messaging, tone of voice, and visual identity added zero revenue. My first landing page was a single Notion doc with a broken English headline. It converted better than the $4,200 branded version I paid a Russian agency to build six months later.
Status meetings. I used to run 11 standing meetings per week. The habit survived the first three months in Panama and produced exactly zero working agents. Deleting every recurring calendar event was the highest-leverage decision of the entire pivot.
The most dangerous useless skill was strategic planning at 12-month horizons. Executives are trained to produce three-year roadmaps with confidence. In multi-agent systems the half-life of any plan is measured in weeks. Groq dropped pricing 48% in one quarter. Claude 3.5 Sonnet replaced 3 Opus two weeks after I finished optimizing for the older model. Any roadmap longer than one sprint became fiction.
The Exact Technical Stack That Emerged From the Pivot
All agents run on Oracle Cloud Always Free tier plus two paid Ampere A1 instances. Total cost $380/month at current volume. The routing layer is 180 lines of Python that decides in 11 milliseconds whether to send a request to Groq Llama-3.1-70B or to Claude 3.5 Sonnet via Anthropic API. The decision tree uses token count, detected language, presence of pricing queries, and current Groq queue depth.
Telegram and WhatsApp agents share the same memory layer built on Oracle Autonomous JSON Database. Each user conversation is stored as a single document with vector embeddings generated by voyage-3. Retrieval latency is 34ms at p95. I stopped using LangChain after it added 240ms of overhead and replaced it with direct SQL queries against the JSON store.
The most reliable agent is the one that books discovery calls. It has a 94% success rate at extracting calendar availability and has booked 187 calls in the last 90 days. The least reliable was the Russian-to-Spanish translation agent that I killed after it hallucinated legal disclaimers three times in one week.
Error handling is deliberately brutal. Any agent that exceeds 2.8 seconds p95 latency gets automatically removed from the routing table for 30 minutes. This has saved the system from cascading failures four times since launch.
Why I Stopped Hiding the Executive Gap
For the first 11 months I presented myself as a “former operator learning AI.” The subtext was apology. I thought technical founders would dismiss the government and corporate background as irrelevant or worse, contaminated.
The data proved the opposite. Inbound leads from developers and technical founders increased 3.1x when I started publishing the exact failure numbers: $11,400 wasted on the over-engineered platform, 19 months to first $10k MRR, 11 production agents running on $380/month Oracle infra.
The gap stopped being a liability the moment I treated it as data. Ex-executives who message me now usually ask the same three questions: how I killed the committees, how I control the monthly burn, and how I handle the identity shift from “person with authority” to “person with working code.”
The answer to the last question is the hardest. Authority was a drug. When I sent an email as Deputy CEO, three departments moved. When I push a git commit at 2am in Panama, nothing happens until the agent is tested in production. The gap between those two states is where the real work occurs.
I no longer hide it because the hiding itself was the last corporate habit I needed to kill. The moment I published the $380 monthly number and the 4,200 daily sessions, the right people started paying attention. They were not looking for another ex-executive with a vision. They were looking for someone who had already shipped under constraints that most founders refuse to accept.
Concrete Numbers From 19 Months of Solo Operation
- Total spend on cloud and APIs: $6,840
- Revenue generated: $41,200
- Hours spent on non-production code: 187 (all in month 4)
- Current agent uptime: 99.94% over last 90 days
- Average time to ship new routing logic: 43 minutes
- Percentage of features cut before building: 67%
- Number of times I almost returned to corporate role: 3 (months 3, 7, and 14)
The third time was the closest. A headhunter offered a Chief Operating Officer position at a Series B infrastructure company with $340k base. I spent 72 hours calculating the exact impact on the agents before declining. The math was simple: the corporate salary would have delayed the next three agents by nine months while adding zero technical capability I did not already possess.
What Developers and Technical Founders Should Take From This
If you are considering an executive career pivot AI developer path, treat your past role as a constraint set, not a credential. The useful parts are the ones that survive contact with production systems that break at 3am. The useless parts are almost everything that required other people to execute.
Start with the budget you can actually afford, not the one you think you deserve. I began with $800 in the bank after relocation costs. That number forced every technical decision that followed. Oracle Always Free tier was not a lifestyle choice. It was the only option that kept the experiment alive.
Measure everything in production sessions, not in GitHub stars or conference talks. The 4,200 daily sessions are the only metric that matters. Everything else is theater.
Stop hiding the gap the moment you have three consecutive months of working code in production. The scar tissue is more valuable than the polished narrative. Technical founders respect the failure numbers more than the previous title.
The pivot is possible. It is also slower, more expensive, and more isolating than the LinkedIn versions suggest. The difference is that once you accept those three constraints, the work becomes simple: ship the next agent, measure the cost per session, reduce it, repeat.