Originally published on AIdeazz — cross-posted here with canonical link.
Frequently Asked Questions
Q: How much runway did you actually burn before the first paid multi-agent deployment?
A: $47k across 11 months. The first $9k MRR came from a Telegram-based procurement agent for a logistics firm in Colombia. Everything before month 9 was pure negative cash flow on Oracle credits, Groq inference, and my own salary.
Q: Which single executive skill moved the needle most when selling to non-technical buyers?
A: Budget variance analysis. I could show a prospect the exact token cost per 1,000 procurement events ($0.018 on Groq LPU + Claude 3.5 fallback) versus their current Excel-and-email process. That closed three deals at $3k–$6k/month each.
Q: Did the Russian Deputy CEO title help or hurt credibility in LatAm tech circles?
A: Hurt for the first six months. I stopped using it entirely after one founder laughed and asked if I was here to “nationalize his startup.” Now the site says “ex-deputy CEO turned solo AI builder.” Conversion rate on cold outreach doubled.
Q: What part of traditional engineering education did you have to backfill at age 41?
A: Production debugging of distributed agents. I could read a 400-page strategy deck in 18 minutes but spent 14 weeks learning how to trace why one agent kept calling the wrong tool 40% of the time on Oracle functions. Real cost: three all-nighters and one $1,200 surprise bill.
Q: How long until the “imposter syndrome gap” stopped affecting pricing discussions?
A: Nine months. The moment I showed a live WhatsApp agent handling 180 supplier queries per day with 94% first-contact resolution, the conversation shifted from “are you even technical” to “how fast can you double capacity.”
The Failure Number That Forced the Pivot
$0 in personal savings after the 2022 sanctions freeze. That was the hard stop. No more deputy CEO salary, no more corporate Amex, no more illusion that the 14-hour strategy sessions were building anything durable. I had a 9-year-old daughter, a half-packed suitcase, and a one-way ticket to Panama where the burn rate was 40% lower than Moscow.
The decision took four days. I killed the corporate email, moved the remaining $11k to a USDC wallet, and told my former board I would not be joining the “parallel import” scheme they were quietly planning. Three weeks later I was in Panama City with a $39/month co-working desk and an Oracle Cloud free tier account.
What Executive Muscle Actually Transferred
Three things survived the translation from deputy CEO to solo AI builder.
First, stakeholder compression. I used to run alignment sessions with 40-person steering committees. Now I compress that into a 9-prompt chain that routes between Groq for speed and Claude 3.5 Sonnet for reasoning. The same muscle that kept a Minister’s aide, three deputy ministers, and a state bank compliance officer on the same page now keeps four specialist agents from hallucinating conflicting supplier data.
Second, variance reporting. Every Friday I still produce a one-page variance report. Only now the columns are “expected token spend,” “actual token spend,” “agent success rate,” and “human escalation rate.” My largest client pays $6,200/month because that report shows them saving $41k in procurement labor. The format is identical to the ones I used to send the Russian government — just the units changed from rubles to dollars and from headcount to inference cost.
Third, ruthless scoping. The average government digital project had 47 success criteria and 9 ministries with veto power. I learned to kill 80% of the scope in the first workshop. That skill now prevents me from building features no one will pay for. My current multi-agent system has exactly four user-facing flows. Everything else is internal routing logic that lives in 180 lines of Python and a Redis cache on Oracle.
What Was Completely Useless
PowerPoint fluency. I can still build a 60-slide deck that would make McKinsey cry, but it has zero correlation with whether an agent survives production.
Committee management. In government you optimize for consensus. In production AI you optimize for latency under 900ms and cost under $0.03 per session. The skills are antagonistic.
Risk registers. The 40-line Excel risk register I once maintained for a $180M national program is useless when your biggest risk is “Claude refuses this prompt at 3 a.m. and the fallback to Llama-3-70B returns JSON that breaks the downstream parser.” The only register that matters now is the one that lives in Sentry and Prometheus.
Russian corporate signaling. The verbal and non-verbal cues that worked in Moscow boardrooms read as arrogance or evasion in LatAm founder circles. I had to relearn how to speak in complete declarative sentences without the protective layers of “as we discussed previously” and “subject to further alignment.”
Why I Stopped Hiding the Gap
Month four was the lowest. I had burned through $29k, had zero paying customers, and was manually copying supplier responses between WhatsApp and Google Sheets for a friend’s trading company just to eat. A technical founder I respected asked me point-blank: “So you were a deputy CEO… what makes you think you can ship production agents?”
I gave the polished answer about transferrable leadership and strategic thinking. He nodded politely and never replied to my next three messages.
Two weeks later I rewrote my entire site. Removed every mention of titles. Added a public dashboard showing live metrics from the three agents I was running for paying customers (even if two of them were $800/month friends-and-family deals). I published the exact routing logic, the exact Oracle function costs, and the exact failure modes.
The gap became the product.
Within 30 days I had two new $2,400/month contracts. Both clients said the same thing: “We were looking for someone who had actually failed in public before asking us to trust them with our procurement data.”
Current Stack and Real Constraints
I run everything on Oracle Cloud Always Free tier plus $180/month of paid compute. Two Ampere A1 instances handle the agent orchestration. Groq LPU instances are the default fast path (mixtral-8x7b and llama3-70b). Claude 3.5 Sonnet is the reasoning fallback triggered when confidence score drops below 0.72.
All customer interaction happens through Telegram and WhatsApp Business API. No web frontend. The agents are multi-turn, maintain 30-day conversation memory in Redis, and can trigger Oracle Functions that call external supplier APIs. Average daily volume: 1,840 messages across all customers. Peak latency: 680ms on Groq path, 4.2s on Claude fallback.
The entire system is 2,400 lines of Python, 41 prompt templates, and one brutally simple retry circuit that falls back to human escalation after three failures. I no longer hide that last part. In fact, I charge a premium for the escalation SLA (under 12 minutes during business hours).
What I Tell Other Executive Career Pivot AI Developers
If you are coming from a senior operating role and trying to become a technical founder, do these three things immediately:
Ship the ugliest possible version in public within 60 days. Mine was a WhatsApp bot that forwarded messages to me after 10 p.m. The transparency built more trust than any polished demo.
Translate one executive skill into a quantifiable AI metric. My variance reports became token-and-resolution dashboards. Pick yours and make it visible.
Price from day one based on the business outcome you create, not your previous title. My first invoice was $900 for a system that saved a logistics company $11k in the first month. The second invoice was $6,200. The title never appeared on either.
The gap is not a liability. It is the only thing that cannot be copied by 26-year-old prompt engineers who have never managed a P&L or stood in front of a board explaining why a $42M project is 11 months late.
I still wake up at 4:47 a.m. some mornings with the old imposter voice whispering that a former deputy CEO has no business debugging async agent deadlocks. Then I open the Oracle console, see the live metrics, and remember the $47k number.
That number bought me the right to stop hiding.
Top comments (0)