MiniAgent — I Trained an Advertising AI for $0.13

Source: YouTube auto-generated captions, lightly cleaned (repeated caption lines removed, misheard brand names corrected).

0:00 — Wasted spend hiding in search

Your last 30 days on this account look ordinary until you split the search terms. Total spend was $2,246. $937 of that went to queries with zero conversions. 73 distinct terms are sitting there as negative keyword candidates. Average CPA is $161.

0:21 — The morning search terms job

A media buyer still has to read every line by hand. That $937 is not a tracking bug. It is search terms that look related and never convert. At $161 cost per acquisition, you cannot afford to learn this after the month closes.

0:42 — A specialist not a chatbot

The work is the same every morning. Open search terms, mark the dead queries, tighten the account before the next dollar goes out. Most teams reach for a giant closed model and hope it understands Google Ads. That model was not trained on your search terms, your negatives, or your CPA. It guesses in prose. You still paste screenshots. The cheaper path is a

1:05 — Step 1 — Train from zero

small advertising agent you train from zero on one GPU in about two hours for 13 cents. Step one is train from zero, not fine-tune a chat model that already has opinions about ads. From zero means the weights start blank and only see advertising tasks. That is why the agent later talks in CPA, match types, and search terms instead of generic marketing copy. You are building a specialist, not a talkative intern. Blank weights force you to choose the data.

You feed search term reports, negative lists, RSA copy, and CPA outcomes. The model learns the pattern that $937 of dead queries share instead of inheriting block advice from the public internet. That is why a two-hour run can beat a frontier model that has never seen your account. The run itself is the part people over complicate. One consumer GPU, about 2 hours, electricity and rent come out to 13 cents. You are not renting a cluster, and you are not waiting on a vendor fine-tune queue. That cost is why you can retrain when the account makes changes, instead of treating the model as a frozen purchase. You still need a floor of signal or the specialist learns noise.

Clean conversion tracking, a search terms export with spend and conversions, a current negative list, so the model sees what you already blocked. That is why you train after the account is readable, not as a substitute for GA4.

2:39 — Step 2 — Wire MCP servers

Garbage in still makes a confident wrong agent. Step two is wiring MCP servers, so the agent can act, not just comment. MCP is how tools show up as functions the model can call. Without them, you have a small brain in a jar. With them, the agent can pull search terms, draft negatives, and read campaign settings. That is why tools come after training, not before. The open build ships 14 MCP servers. Think in jobs, not in vendor names.

Search terms, negatives, RSA assets, budget and bid reads, conversion checks, change logs. 14 is enough coverage for a media buyer desk, and small enough that you can audit every tool. That is why the number is 14, not 100 wrappers you will never inspect. On this account, the first useful call is pull search terms with spend and zero conversions. The agent sees $937 across 73 terms, and proposes negatives instead of writing an essay.

That is why MCP matters. The model was trained to think in advertising objects, and the servers let it touch those

3:52 — Step 3 — Keep it open source

objects without you pasting a CSV into a chat window. Step three is keep the whole stack open source. Weights, training recipe, and the 14 MCP service. Open source is not a slogan here. It is how you confirm the agent is not quietly calling a vendor model with your account data. That is why you can run it on a machine you own after the $0.13 training pass. Closed advertising AI means you cannot read the prompt, the tool list, or the fine-tune data.

Mini agent is the opposite. You can diff the recipe, drop a server, or retrain on last 30 days when $937 shows up again. Ownership is why a 2-hour local run is worth more than a polished demo that you

4:39 — Why this order matters

cannot audit. This order is the whole method. Train from zero, so the brain only knows ads. Then attach 14 MCP servers, so that brain can touch campaigns. Then keep it open, so you can inspect both. Reverse it, and you bolt tools onto a generic chatbot, or you train a specialist that cannot act. Sequence is the product, not a setup preference. Once it is running,

5:04 — What you actually ask

you ask it like a senior sitting next to you. Show search terms with spend and zero conversions. Draft 73 negatives. Explain why CPA is $161 on this match type. The answers stay short because the training data was reports, not blog posts. That is why it feels like a media buyer, not a copywriter. Buddy by Ahmeego is the full

5:27 — Where Buddy fits

advertising agent you run in production. Mini agent is the teaching stack. Same desk, different weight. You use Mini agent to learn how a specialist is trained, how MCP tools attach, and why open weights matter. Then you let Buddy carry the live account, so you are not babysitting a 2-hour research model. Put the live account next to the training

5:50 — Waste versus training cost

bill. $2,246 spend, $937 to zero conversion terms, 73 negatives waiting, $161 CPA. Against that, a specialist trained from zero in 2 hours on one GPU for 13 cents. The point is not that 13 cents replaces a media buyer. The point is the buyer finally has an agent that speaks

6:16 — Recap

the same objects. Train from zero so the model only sees advertising tasks. Wire 14 MCP servers so it can read search terms and draft negatives. Keep the recipe open source so you can audit weights and retrain for 13 cents. That order is why a 2-hour GPU run becomes a desk tool instead of a demo you cannot inspect or rerun. If $937 is dying in search terms this month, train the small agent first and read what it proposes.

2 hours, one GPU, $0.13, 14 MCP servers, open source from zero. The write-up with the recipe is on the blog. The blog is on ahmeego.com. Buddy by Ahmeego, a company of It All Started With an Idea.