The AI Operator Stack
The exact systems running my consulting business today. Everything here is live and doing real work, not a demo. I write about how each one works, with numbers, in the newsletter.
01
The brain
One AI agent (I call it my AI partner) that lives in a single repo with every script, playbook, and client file. It reads the business context, writes and runs its own code, and remembers decisions between sessions through a persistent memory layer.
Tools: Claude Code, a git repo as the operating system, Supabase for long term memory
Result: Every system below was built and is operated through this one interface.
02
Ad ops autopilot
Twice a day, an agent audits every client ad account against per client rules: cost per lead thresholds, zero lead kill rules, budget ceilings. It sends one consolidated scorecard to my phone, auto pauses unambiguous losers after a veto window, and stages scaling proposals I approve with one tap. Hard budget caps it can never cross, plus kill switches that only I can trigger.
Tools: Meta Graph API, Google Ads API, Python on a small VPS, Telegram for the scorecard and controls
Result: Daily account management happens without me. I make only the judgment calls.
03
Creative engine
A pipeline that writes ad scripts, generates UGC style talking head video, runs every clip through a QA gate (speech pacing, lip sync, audio levels, compliance against the brand brief), and launches survivors into a live test ad set, rotating out the oldest ad. Fails closed: anything questionable is held for my review.
Tools: Gemini for scripts and multimodal QA, Seedance via fal.ai for video (about $1.45 per clip), Whisper for transcription, Remotion for assembling long form cuts
Result: Fresh tested creative every week without an editor or an agency.
04
Voice ops
Every coding session speaks a short summary of what it just did through a local text to speech daemon. A voice relay listens for my dictation, cleans it up with a small local model, and routes it to whichever session spoke last. All local, zero cloud cost per turn.
Tools: faster-whisper for speech to text, a small local LLM for cleanup, Kokoro for text to speech, macOS launchd
Result: I operate multiple parallel work sessions hands free while doing other things.
05
Lead generation
Scrapes local business listings by city and category, deduplicates against everything ever pulled, scores each prospect 0 to 100 on rating, reviews, photos, verification, and hours, then exports a call priority list.
Tools: Outscraper on Google Maps data, Python scoring pipeline, CSV out
Result: A ranked call sheet instead of a raw list. Score 80 plus means call now.
06
Competitive intel
Pulls competitor ads from the Meta Ad Library on a schedule and flags the winners: ads that stay active for weeks are ads that make money. Same approach for TikTok.
Tools: Apify scrapers, Meta Ad Library, Python winning ad detection
Result: I copy what provably works before writing anything new.
07
Finance
Imports transactions from three banks in three currencies, auto categorizes them with direction aware rules, excludes internal transfers, and serves one dashboard with the real numbers.
Tools: Bank APIs and exports, Python, Postgres, a small web dashboard
Result: Books that are always current. Zero bookkeeping hours.
08
Client comms
A Telegram bot that knows the whole book of business. It answers questions about any client or deal in plain English, reads the ad scorecards, and stages any money action (pause, scale, generate creative) behind a one tap approval that only I can give.
Tools: Telegram Bot API, Gemini function calling, the same memory layer as the brain
Result: I run the operation from my phone in plain English.
Watch it get built
Every week I publish one system like these: what it does, what it cost, what it returned.