What is AI Engineering: A Practical Guide for Solo Operators

AI engineering isn’t a fancy title for data scientists; it’s the practice of turning AI models into repeatable, revenue‑generating workflows you can run yourself.

After reading this, you’ll know what components make up an AI engineering system, see a real cold‑email pipeline built with GPT‑4 and Make.com, and decide whether to assemble it from scratch or pull a pre‑built blueprint from /vault.

What most guides get wrong about AI engineering

Most articles treat AI engineering like a magic button you plug into a CRM and forget. They talk about “agents” and “chains” as if the model itself does all the work. In reality the hard part is the glue: data formatting, error handling, and the loop that decides when to re‑prompt or fallback. If you skip those pieces you end up with a demo that works once and then dies on real input.

One‑sentence truth: the model is only as good as the pipeline that feeds it and cleans up after it.

How do you actually start building an AI workflow?

Begin with a single, narrow goal. For a solo operator that goal is often lead generation: take a list of prospects, write a personalized first line, and send it via email. You don’t need a fleet of agents; you need three blocks: a data source, a prompt engine, and a delivery method.

First, pull your prospect list from a Google Sheet or CSV. Second, feed each row into a prompt that asks the model to produce a short, friendly opener based on the prospect’s role and recent news. Third, use an automation tool to take the model’s output and drop it into an email draft or send it directly via SMTP.

Keep the loop tight: run the prompt, check the output for obvious flaws (like missing name or hallucinated company), then move to the next step. If the output fails a simple sanity check, retry with a tighter prompt or skip the record.

A concrete named example: cold‑email pipeline with GPT‑4 and Make.com

Here’s how I built a working version last month. I used Google Sheets as the source, GPT‑4 via the OpenAI API for the copy, and Make.com (formerly Integromat) to orchestrate the steps.

1. Create a sheet with columns: Name, Email, Role, Company, NewsSnippet.
2. In Make, add a “Watch Rows” module that triggers when a new row is added.
3. Add an “HTTP > Make a request” module to call the OpenAI chat/completions endpoint. The prompt I used:
   You are a friendly sales assistant. Write a one‑sentence opening line for a cold email to {{Name}} who works as a {{Role}} at {{Company}}. Reference this news snippet: {{NewsSnippet}}. Keep it under 20 words, no fluff.
4. Parse the JSON response, extract the generated text, and map it to a new column called “IntroLine”.
5. Add a “Filter” module that checks if IntroLine is empty or contains the word “sorry”. If it fails, send a Slack alert to me and stop the row.
6. Finally, use the “Email > Send an email” module to draft a message in Gmail: subject line “Quick question about {{Company}}”, body: “Hi {{Name}},\n\n{{IntroLine}}\n\nLet me know if you’d like to see how we help companies like {{Company}} reduce churn.\n\nBest,\n[Your Name]”. Set the email to stay in Drafts so I can review before sending.

The whole scenario runs in under two seconds per row. I’ve processed 500 leads a week with a manual glance at each draft before hitting send.

Cost breakdown: Google Sheets is free, Make.com’s core plan is $29/mo (which, yes, is annoying if you need more than 10k operations, but fine for solo work), and GPT‑4 usage was about $8 for 500 rows at $0.03 per 1k tokens (roughly 250 tokens per call). Total monthly spend stayed under $40.

How to debug when this breaks

When the pipeline stalls, the first place to look is the Make scenario log. Each module shows its input and output; a red flag usually means the HTTP step returned a non‑200 status or the JSON parser couldn’t find the “choices” field.

If the OpenAI call fails, check your API key and quota. A 429 means you hit the rate limit—pause the scenario for a minute or switch to the cheaper gpt‑3.5‑turbo for the first pass and only use GPT‑4 for rows that need extra nuance.

When the IntroLine comes back empty or weird, the prompt is likely too vague. Tighten it by adding explicit constraints: “Do not mention price, do not ask for a meeting, do not exceed 20 words.” Re‑test with a single row in the OpenAI playground before pushing to Make.

If the email draft lands in spam, verify that your Gmail account has proper SPF/DKIM and that the body isn’t just a single line. Adding a short plain‑text fallback helps.

Finally, keep a simple error table in a second sheet: timestamp, row ID, error module, message. Review it weekly to spot patterns like a particular news snippet that consistently confuses the model.

Pricing opinion: what’s fair and what’s ridiculous

$29/mo for Make.com’s core plan is fair for the automation you get—enough operations for a few hundred leads and basic error handling. Jumping to the $108/mo team plan only makes sense if you’re running dozens of parallel scenarios or need premium apps like Salesforce.

I think $199/mo for a dedicated “AI agent builder” that promises “no‑code AI” is ridiculous for what you get. Those platforms lock you into a proprietary runtime, charge per execution, and still require you to write prompts and handle data yourself. You end up paying a premium for a drag‑and‑drop canvas that you could replicate with Make and a few HTTP modules for a fraction of the cost.

The free tier of Make.com is enough for solo work if you stay under 1k operations a month and don’t need premium apps. Beyond that, the paid plan is worth it.

Love and gripe: real tools I use

Love: GPT‑4’s ability to follow multi‑step instructions when you give it clear boundaries. In the cold‑email pipeline it consistently produced a usable opener that needed only a light edit.

Gripe: Make.com’s error logging is cryptic—when a module fails you get a generic “Invalid input” message with no clue which field caused it. I’ve wasted hours digging through the raw payload just to find a missing comma in a JSON string. (which, yes, is annoying)

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.