AI CMO OS

Proof

Built inside real companies before it was a product.

Numbers are given as ratios and timings. Client revenue, databases and infrastructure stay with the clients; the method and the code came with the operator.

Crypto and forex trading platform · 2024–2025

PrimeXBT

Situation

VP Growth & Partnerships role: a $7M quarterly budget across paid, partnerships and regional teams, a 30+ person global team, and an influencer and introducing-broker program that lived in spreadsheets and inboxes.

What ran

  • KOL and partner sourcing moved into a pipeline: scrape → enrich (audience, engagement, contacts) → score → CRM → sequenced outreach, with every touch logged against the lead
  • Web funnels and landing pages by geo, offer and channel, tied to performance data and A/B-tested continuously
  • KPI framework rebuilt around pipeline stages instead of vanity reach
  • Regional teams, BDs and content worked from the same CRM and the same lead states

Outcome

  • Hours per processed lead down 40%, measured before and after the pipeline
  • More processed leads, more closed deals, more volume and revenue per period on the same team
  • Budget held at up to 120% ROI, 90–120% on the top channels
  • The method became the outreach and KOL modules of the OS

Tech conference · 2026

Tarija Tech Week

Situation

A three-day technology conference in Bolivia with a small organizing team, weeks before the event and no sponsor pipeline in a system.

What ran

  • Outreach module: CRM in Sheets, D0/D1/D3/D5 multichannel sequences (email threads and WhatsApp), reply watcher, KB auto-reply with escalation
  • Deliverability module: SPF, DKIM, DMARC checked before the first send
  • Events kit: speaker funnel with payment on the site, listings, press kit, agenda PDF regenerated automatically, llms.txt and search console
  • Site and app kept in sync from one database; an AI voice agent answered questions about the agenda

Outcome

  • 48 hours from zero to 200+ warm sponsor leads in sequences
  • 81 WhatsApp first touches and 159 emails on day one; about 30% of the purchased email list bounced, which the system reported instead of hiding
  • First replies within hours; follow-ups stopped automatically on reply
  • The WhatsApp sales agent closed two sponsorship deals in the final days before the event; Binance came in as the main sponsor
  • Lesson applied to the product: first touches at scale go through official channels, never a personal linked device

Consumer AI product · 2026

Adult AI companion app

Situation

A subscription product with paid traffic from several ad networks, a drop in marketing profit and no trustworthy LTV number.

What ran

  • Analytics: LTV rebuilt as survival-weighted revenue per acquisition; marketing profit formula agreed with the founder; holistic sheet with CAC by source
  • Paid: ad-network monitor, source fraud detector, RTB control loop with human approval
  • Retention: in-house drips on Mailgun and Mailjet, segmentation, warm-up and deliverability recovery after a domain burn
  • Payment watcher for reject spikes; SEO tag machine; ORM and listings

Outcome

  • A fraudulent source was caught and cut before the budget cycle ended
  • The profit decomposition showed the drop was CAC, not LTV, and the paid plan changed accordingly
  • Deliverability recovered with an apology flow instead of a new domain burn
  • The founder reviews one formula every month: subs × (0.8 × LTV − CAC) − salary

Telegram AI aggregator · 2026

30M+ user AI bot

Situation

A large consumer bot with many acquisition channels, a small marketing team and attribution that reported cost per action instead of revenue.

What ran

  • Outreach machines: KOL drip, forum drip, listings research, with Monday as CRM and auto-stop on reply
  • SEO machine in several languages with an approval bot so the team signs off from Telegram
  • Telegram Ads: attribution moved to revenue per diamond, auto-budget with a kill switch
  • Launch framework for new models and products; manager KPIs; quest economics

Outcome

  • Attribution now reflects revenue, so budgets move toward sources that pay back
  • Three outreach machines run weekly with one person reviewing replies
  • Content ships on a schedule in multiple languages without a writer on payroll
  • Every launch uses the same checklist and the same board

Blockchain conference · 2025

Bolivia Blockchain Conference

Situation

Founding and organizing the country's largest blockchain event from zero: speakers, sponsors, ticketing, app and live stream, with a small team.

What ran

  • Direct industry BD for the full speaker and sponsor roster
  • In-house event app with local payment rails and in-app live streaming, shipped to App Store and Google Play
  • Outreach, press and listings from the same playbook later used at Tarija Tech Week

Outcome

  • 3,000 attendees on site and 1,000 online
  • 100-speaker lineup and the full sponsor roster closed through direct BD
  • Event app reached #2 in Education on the App Store in Bolivia, ahead of Duolingo at #4

Machine by machine

Everything that was built and run across these deployments, grouped by module. Each entry says what it does and what it changed.

Paid media and RTB

Real-time rules engine for ExoClick

Every 15 minutes: live stats from the network API, purchases from the warehouse, rules for pause, resume-winner, budget and zone actions, 20-hour cooldown per action, execution log posted to Discord with before/after.

Result: Switched to live apply by the founder; other networks kept in recommend-only mode until proven.

Telegram Ads RTB control with approval

A scanner reads the ad account every 30 minutes and queues proposals (top-up, stop, new ad, reactivate); the operator gets a Telegram card with approve / decline / edit; only approved items execute. Thresholds live in a sheet, not in code.

Result: Nothing touches money without a button press. Replaced third-party budget tools that required access to the account's funds.

Telegram Ads attribution fix

Campaign tags joined to payers and revenue instead of cost per bot start.

Result: The cheapest-CPA campaign turned out to be the worst money (0.4% conversion to payment). Budget moved to the tags that actually paid back.

Meta alerts across channels

Total sales drop vs 7-day median, cross-channel conversion break, per-channel visit-to-pay alerts.

Result: First trigger caught four channels falling 41–76% on the same day: a checkout problem, not a campaign problem.

Ad-network balance monitor

Every 15 minutes, balance and runway per network, alerts to Discord before a campaign stops for lack of funds.

Result: Runs as a system service since May 2026.

Media-buyer board

Campaign CAC, zone blacklist, publisher domains, daily trend and reconciliation of network sales vs warehouse sales; a Discord digest with cut candidates.

Result: Replaced a buyer's daily spreadsheet ritual with one board and one morning message.

Source fraud detector

Tests engagement after payment, not conversion: real subscribers chat ~88% of the time, bought ones pay once and go silent.

Result: A source with 47% signup-to-deposit and 9% post-payment engagement was cut before the budget cycle ended.

Outreach, KOLs and partnerships

KOL outreach stack (adult AI product)

Lead batches, email enrichment across 17 contact-page patterns (39% of leads without email recovered), Kommo pipeline, D0 send, follow-ups at 7 and 14 days, inbox watcher every 30 minutes that moves replies to the right stage and pings the team.

Result: 65 of 65 emailed leads processed in one session; the same code was reused for a partner pipeline with a 0/+2/+4 day cadence.

KOL discovery for a Telegram AI bot

Nightly YouTube discovery across ~24 audience niches, metrics and emails enriched, deduplicated into a Monday board, rate-limit aware.

Result: The board fills itself overnight; the manager starts the day with scored leads.

Outreach drip for a manager's laptop

First touch by hand after reviewing the lead, then D+1 / D+3 / D+7 automatically with auto-stop on reply; packaged to run on the manager's Mac.

Result: One person runs weekly KOL outreach without a server.

Sponsor outreach for a conference

200+ warm leads, multichannel D0/D1/D3/D5, knowledge-base auto-reply with escalation, 81 WhatsApp + 159 emails on day one.

Result: First replies within hours; two sponsorship deals closed over WhatsApp in the final days before the event, with Binance as the main sponsor. About 30% of a purchased list bounced and the system reported it.

Forum warm-up engine

Twelve forums, two phases: two weeks of on-topic expert replies, then one native thread a week with a soft deep link; typo humanizer; session reuse where captchas block automation.

Result: Dry-run by default; nothing posts live without the flag.

Listings and directory research

Finds and qualifies directories, forums, Reddit communities and 'top tools' listicles in RU and EN markets, filters by traffic and topic, loads to Monday.

Result: Two research machines feed one outreach drip.

Retention and email

In-house behavioral drips

Mailgun plus a second Mailjet pipe on a separate subdomain, segmentation by cohort, language derived from geography (es, pt, de, fr, it), suppression and consent handling.

Result: Around 9,200 behavioral emails a day measured against a holdout group.

Domain burn recovery

A cold wave burned a fresh subdomain (open rate under 2%, 8% bounce). Diagnosis, apology flow to paying customers, warm-up schedule, second pipe.

Result: Inbox placement recovered without burning another domain; the rule 'no send without SPF, DKIM, DMARC' became hard.

Payment reject watcher

Hourly rolling two-hour window against a same-time-of-day 14-day baseline; alerts only on attempts, rejects, ratio and absolute thresholds together; four-hour cooldown.

Result: Answers 'is it payments or is it traffic?' in one message.

Analytics and money

Survival-weighted LTV and marketing profit

LTV from aging triangles that exclude censored cohorts; profit = subs × (0.8 × LTV − CAC) − salary agreed with the founder; holistic sheet with CAC by source and by manager.

Result: A monthly profit drop was traced to CAC, not LTV; the paid plan changed instead of the product.

LTV calculator on 38M registrations

Full export of a Telegram bot's history (38.0M registrations, 386k payers) turned into a live sheet: LTV curve, unit economics, CAC-to-payback by channel.

Result: Channel revenue-share deals priced from LTV per registration instead of guesses.

Managers performance board

Manager-to-partner mapping joined to warehouse revenue; one Redash board per team.

Result: Weekly review takes a link, not a deck.

Content, SEO and owned media

SEO tag machine (adult AI product)

Ten tag pages a week from a tracked keyword universe, strict format, FAQ schema, published through the site API.

Result: Organic pages ship on a schedule without a writer.

AutoSEO for a presentation tool

Keywords from live search suggestions in seven markets, articles in en/es/pt with hreflang, Ghost publishing, a Telegram approval bot for the product owner.

Result: First nine articles across three languages went live in a day.

This blog

The same pipeline, pointed at marketing operations keywords, publishing weekly in English and Spanish.

Result: You are reading the output.

Owned media newsroom (Latin America)

Nine Telegram channels, RSS listening every three minutes, LLM pick of what is worth posting, house voice, day/night cadence by local time zone, cross-promo digest of the best post per country.

Result: Runs as a daemon; alerts the operator after ten consecutive failures.

Creative factory

Prompt library, ComfyUI workflows (image and image-to-video), assembly scripts for ad formats, localization of winning videos into five languages with voice and captions.

Result: Packaged as a portable kit that sets itself up from a few answers.

Landing studio

Eight landing types built from section schemas through an API: scroll, quiz, paywall, scenario, registration and more.

Result: A brief becomes a published landing without a designer or a developer.

Interactive scenario landings

Branching point-of-view video with choices and a signup gate at the peak; scenes generated and hosted, published through the landing studio, auto-translated.

Result: The reference scenario converted visit-to-signup at 2.56%, the best landing type in the account.

Promo video content factory

Product promo clips assembled by script: title cards, in-screen demos, chat mockups, music track, portrait 1080×1350, versions per language.

Result: A 31-second launch video in two languages from one build command.

Banner studio

Marketing banners, flyers and social posts as HTML/CSS rendered to PNG with a brand kit, so text stays sharp; optional AI hero background underneath.

Result: Any size, any language, no designer in the loop.

Community and product marketing

Discord insiders program

A gated section for feedback authors on a 1,400-member server, role-based access, invite flow from email.

Result: Customer care gets feedback directly instead of through tickets.

Reddit seeding, managed

Operating model with a managed provider: briefs, templates and tone from us, posting operations from them, monitoring on their portal.

Result: Zero internal headcount on Reddit.

Launch framework

A master checklist of 53 tasks, 10 pillars × 5 phases, instantiated per product, feature or new region; payments flagged as the usual blocker.

Result: Every launch runs on the same board.

Community from zero (Super Protocol)

As CMO of a DePIN cloud project: community program, testnet campaign, ambassador program through university partners on three continents.

Result: 0 to 50,000+ community members and 10,000 testnet participants.

Product and event technology

Event platform: site + app in sync

Conference site generated from one database (speakers, sessions, venues, sponsors), synced to the mobile app through a signed API, agenda PDF regenerated on every change, side events with access rules, OG images per panel.

Result: 117 speakers and 120 sessions kept consistent across site, app and PDF for three days of Tarija Tech Week. The same approach at Bolivia Blockchain Conference: site plus an event app with local payments and in-app live stream, 3,000 people on site and 1,000 online, app #2 in Education on the Bolivian App Store, ahead of Duolingo.

Voice AI agent for an event

Browser speech in, LLM with retrieval over the event database, synthesized voice out; falls back to a second model when the first one rate-limits; knows whether the event is upcoming, live or over.

Result: Attendees asked about agenda, speakers and tickets by voice from the home page.

WhatsApp speaker bot and Cloud API path

Speaker-facing WhatsApp assistant on the official Cloud API, after a personal linked device was banned for bulk sends.

Result: The ban became a hard rule in the OS: first touches at scale go through official channels.

Catalog-to-WhatsApp storefront

A data-driven catalog with cart, delivery zones and payment method that sends a structured order to the merchant's WhatsApp; built from a CSV and a JSON per merchant.

Result: A small business goes live in days without learning any software.

Want the same kind of notes about your company in 90 days? Start the setup.