6 AI automation examples,
all running in production

Example 1

An AI media-buying autopilot with seven agents and circuit-breakers

A German marketing agency manages high five-to-six-figure monthly ad spend for its clients. The autopilot I built for them syncs ad-account data nightly, analyzes the performance of every active ad, makes kill or scale recommendations, and generates challenger creatives to test against the incumbents. Technically, it runs as seven orchestrated AI agents inside the agency's custom operations platform.

Because software that touches ad budgets can do real damage, the system is fenced: budget circuit-breakers cap what it can change, and the rollout was staged, with a dry-run mode first so it recommended before it acted. 104/104 tests were green at handover of the build. The working first version was reviewed the same evening the plan was approved, then hardened over the following three days with an admin control panel. Full case study: client operations platform.

Example 2

An attribution fix: 1 tracked lead corrected to 355

Not every high-leverage automation is glamorous. One of the same agency's client accounts showed exactly 1 tracked lead against real ad spend; by the numbers, the ads were dead. The actual problem was measurement: lead events were not mapped to the right custom conversions for that ad account.

The fix, per-account custom-conversion mapping through Meta's APIs, corrected the account from 1 tracked lead to 355: the difference between "ads are dead" and a working funnel. Diagnosed one day, live in production the next. It belongs on this list because every AI system above and below it depends on this layer: agents making kill or scale decisions on top of wrong attribution data would confidently do the wrong thing. It ships as part of the same platform.

Example 3

An 863,484-contact database searched in plain English

A US B2B sales-intelligence company had its contacts scattered across 20 export files in two incompatible formats. I consolidated them into one Supabase database: 863,484 contacts after deduplication, with zero duplicates remaining, 1.2M+ previously-empty fields enriched from the source exports, and 8/8 QA checks passed before handover.

The AI part is the interface. A custom GPT connected to Supabase RPC functions lets the client's whole team search and count the database in plain English, no SQL and no data person required. The pattern generalizes: clean deterministic data underneath, a natural-language layer on top. Full case study: contact-data engine.

Example 4

Supplier-invoice OCR with a 0.03 DWT gate and live gold prices

A US fine-jewelry retailer, 23 years in business, was checking scanned supplier casting invoices by hand: order numbers, weights, and metal rates typed into sheets, with real money lost to unnoticed discrepancies. The pipeline that replaced this reads each scanned invoice with Google Cloud Vision OCR, flags any weight discrepancy over 0.03 DWT, cross-checks billed metal rates against live gold and platinum spot prices, and updates the company's sheets automatically.

The design principle here is worth stealing: AI reads, deterministic rules decide. The 0.03 DWT threshold is a hard gate, not a model's opinion, so a human only gets involved where a discrepancy actually exists. The pipeline is part of an internal operations platform of roughly 25,000 lines of code across 14 service modules. Full case study: e-commerce operations suite.

Example 5

A work-order pipeline: email in, 28-column CSV out, FTPS delivery

A US facilities-services company was re-typing every incoming work order from a client platform into their field-service system, by hand, every day. The pipeline I built parses each order from email, converts it into the exact 28-column CSV their field-service system requires, uploads it via FTPS, polls for success or error, and logs every order. It replaced daily manual re-entry and runs unattended.

Honesty note: this pipeline dates from the n8n phase of my practice and is deliberately deterministic, not an LLM. I include it because a credible automation portfolio mixes AI where judgment is needed with plain reliable code where it is not, and because operating systems like this is what moved me fully to custom code. Full case study: work-order automation.

Example 6

AI creative generation that publishes directly to Meta

Back inside the agency platform: an AI layer that generates ad copy and creative images and publishes them directly to Meta, not into a folder someone has to upload from. The agents behind it run on Anthropic Claude and OpenAI models, in production, for a business that lives on ad performance.

Context for scale: the platform around it spans roughly 130,000 lines of TypeScript, 229 API routes, and 25 scheduled background jobs, designed, built, and operated by one person. The related campaign-creation engine, with placements, audience presets, and multi-video upload with automatic Whisper transcripts, was built and tested, 54/54 green, within a day of the client call that requested it. Full case study: client operations platform.

What the six have in common

Three patterns repeat. AI sits where judgment is needed; deterministic code sits where correctness is needed; and hard gates sit everywhere money moves, from budget circuit-breakers to a 0.03 DWT threshold. As for the build speed behind them: I orchestrate AI agents for implementation while owning architecture, review, and verification personally. That workflow is why systems like these ship in days rather than quarters, and why each one still has a test suite.