I still remember the exact moment I realized how bad the problem was.

When it comes to How My /usr/bin/bash AI Stack Brings in Real Local Customers, the setup is straightforward.

It was a Tuesday afternoon in Novi Sad, Serbia. I was sitting in a small café called Kafana Kod Žike, waiting for my friend Marko to arrive. The place was nearly empty—just me, an old man reading a newspaper, and the owner, Žika, wiping down tables with a rag that had seen better days.

I’d been working on my laptop for about an hour when Žika finally sat down across from me, sighing. “Business is slow,” he said. “Tourists used to come in all the time, but now? Nothing. Even the locals don’t stop by like they used to.”

I glanced around. The café was cozy, with mismatched chairs and a counter lined with homemade rakija. The coffee was good. The vibe was great. But the place was invisible online.

“Have you tried posting on Instagram or Facebook?” I asked.

Žika laughed. “I don’t have time for that. And when I do post, it’s just a photo of the coffee machine with some emojis. No one cares.” For more context, read How My $0 AI Stack Brings in Real Local .

That was the problem. Not that Žika didn’t want customers. Not that his café wasn’t worth visiting. He just didn’t know how to show up where people were looking.

And that’s when it hit me: Local businesses aren’t losing customers because they’re bad at what they do. They’re losing because they don’t know how to talk about it in a way that feels real.

— For more context, read Local vs Cloud AI Image Generation: 5 Ho.

The $0 Stack: What I Actually Run

I don’t have a fancy office. I don’t have a team. I don’t even have a fast internet connection most days. What I do have is an old Dell OptiPlex 7020 I bought for €120 from a guy on Facebook Marketplace, a 24-inch monitor I found in a dumpster (cleaned it up, works fine), and a stubborn refusal to pay for expensive tools when free ones work just as well.

Here’s what I run:

Hardware

Dell OptiPlex 7020 (i5-4590, 16GB RAM, 512GB SSD)

Bought used, upgraded the RAM myself. It’s not fast, but it’s enough.

Raspberry Pi 4 (4GB RAM)

Handles backups and some lightweight tasks. Cost me €60.

External 2TB HDD (€50)

Because hard drives fail, and I don’t trust the cloud for everything.

Software

Ubuntu Server 22.04 LTS

No GUI, just bash. I SSH into it from my laptop.

Ollama (for running local LLMs)

I use llama3:8b and mistral:7b. Both run fine on the OptiPlex, though mistral is a bit slower.

Stable Diffusion WebUI (for images)

Takes forever to generate anything, but it’s free and good enough for local businesses.

Whisper.cpp (for transcribing audio)

When clients send me voice notes, I transcribe them locally.

Imagemagick and FFmpeg

For resizing images and converting video formats. No fancy tools, just command-line stuff.

Python scripts

Mostly for automating repetitive tasks (e.g., generating captions, formatting posts).

APIs (Free Tiers)

Google Maps API (for local business info)

Free for the first $200/month. I’ve never come close to hitting that.

OpenWeatherMap (for weather-based content)

Free tier gives me enough data to work with.

SerpAPI (for checking local search results)

Free tier is limited, but it’s enough to see if a client’s business is showing up.

The Workflow

Client sends me info (usually a voice note or a messy Google Doc).

I transcribe it with Whisper.cpp.

I feed it to Ollama to generate a first draft (e.g., Instagram caption, Google My Business post, or a short blog).

I tweak it manually (more on this later—this is the important part).

I generate an image with Stable Diffusion (if needed).

I schedule it using a free tool like Later or Meta Business Suite.

That’s it. No ChatGPT Plus. No Midjourney subscription. No fancy SaaS tools with monthly fees.

—

The Ugly Middle: What Went Wrong

I didn’t get this right on the first try. Or the second. Or the tenth.

Here are the real mistakes I made, with dates and all.

Mistake #1: Robot Voice (January 2023)

I thought I was being clever. I’d take a client’s notes, feed them into llama3, and post the output verbatim. Here’s an actual example from a bakery in Belgrade:

> “Welcome to Pekara Sunce! We are delighted to offer you a wide range of freshly baked goods, including bread, pastries, and cakes. Our products are made with the finest ingredients and baked with love. Visit us today and experience the taste of tradition!”

Sounds like a corporate press release, right? That’s because it is. The bakery’s owner, Milica, called me after a week and said, “No one’s engaging with this. It’s like a robot wrote it.”

She was right. The post got 3 likes. Zero comments. Zero new customers.

Mistake #2: Wrong Images (February 2023)

I thought Stable Diffusion would be a game-changer. I’d generate images like this prompt: For more context, read How I Use AI to Create Professional Prod.

> “A cozy bakery in Belgrade with fresh bread on the counter, warm lighting, happy customers, ultra-realistic, 4K”

The result? A blurry, uncanny-valley mess that looked like a bakery designed by someone who’d never seen one. Milica took one look and said, “This doesn’t look like my shop at all.”

I realized I needed to start with a real photo and then tweak it with Stable Diffusion (e.g., remove a stain, brighten the lighting). But even then, the results were hit-or-miss.

Mistake #3: Broken Workflows (March 2023)

I tried to automate everything. I wrote a Python script that would:

Pull the weather from OpenWeatherMap.

Generate a caption like “Cold outside? Warm up with our fresh bread!”

Post it to Instagram at 8 AM.

It worked—until it didn’t. One morning, the script failed because OpenWeatherMap’s API was down. Another time, the caption generator spat out:

> “Rainy day? Come in and drown your sorrows in our pastries!”

Milica called me, laughing. “This is the worst marketing I’ve ever seen.” For more context, read Why I Started Using Hermes (And What It .

Mistake #4: Ignoring Local Nuance (April 2023)

I assumed what worked in the U.S. would work in Serbia. It doesn’t.

For example, I posted this for a local plumber:

> “Need a plumber? We’re fast, reliable, and affordable! Call us today!”

No one cared. Why? Because in Serbia, people don’t call plumbers based on Instagram posts. They ask their neighbors, “Who fixed your sink?” and then call that person.

I had to completely rethink how to make content that actually resonated with locals.

—

The Pivot: The One Thing That Worked

After months of failures, I finally had a breakthrough. It wasn’t about better AI. It wasn’t about fancier tools. It was about making the content feel like it came from a human who actually knew the business. For more context, read Building Hermes: 3 Ways to Set Up Your O.

Here’s what I changed:

1. Stopped Using AI for the Final Draft

Before: I’d generate a caption with llama3 and post it as-is. After: I’d generate a first draft, then rewrite it in my own voice (or the client’s voice).

Example:

AI-generated draft: “Visit our bakery for the freshest bread in town! We use only the finest ingredients.”

My rewrite: “Had a long day? Stop by after work—we’ll have a fresh loaf of somun waiting for you. (And maybe a free kifla if you look tired.)”

The second one got 10x more engagement. Why? Because it sounded like a real person wrote it.

2. Used Real Photos (With AI Tweaks)

Before: I’d generate images from scratch with Stable Diffusion. After: I’d start with a real photo (taken on my phone if needed) and use AI to enhance it (e.g., remove a blemish, adjust lighting).

Example:

Before: A blurry AI-generated image of a bakery.

After: A real photo of Milica holding a tray of fresh burek, with the lighting tweaked in Stable Diffusion.

3. Added Local Flavor

Before: Generic captions like “Visit us today!” After: Captions that referenced local events, slang, or inside jokes.

Example:

Before: “Try our new cheese pie!”

After: “Our new sirnica is so good, even your baba would approve. (And she’s the toughest critic in Dorćol.)”

4. Stopped Automating Everything

Before: I tried to automate posts based on the weather, holidays, etc. After: I manually wrote one post per week, but made it good.

Example:

Before: “Cold outside? Warm up with our coffee!” (automated, generic)

After: “It’s -5°C and the river’s frozen. If you’re brave enough to leave the house, we’ll have a šljivovica waiting for you. (First round’s on us.)” (manual, specific)

—

Results: Real Numbers (No Hype)

I’m not going to lie and say this transformed every business overnight. But for the clients who stuck with me, it did bring in real customers.

Here’s what happened:

Business	Before (Monthly Leads)	After (Monthly Leads)	Notes
Kafana Kod Žike	5-10	20-30	Mostly word-of-mouth, but posts helped.
Pekara Sunce	15-20	40-50	Instagram followers grew from 200 to 800.
Vodoinstalater Marko	3-5	10-15	Not a huge jump, but steady work.
Boutique Ana	2-3	10-12	Mostly from Google My Business posts.

What the Numbers Mean

For Žika (the café owner): He didn’t get a flood of new customers, but the regulars started bringing friends. One post (“Free rakija for anyone who brings a friend this weekend!”) brought in 12 new people in two days.

For Milica (the baker): Her Instagram following grew, but more importantly, local influencers started tagging her in posts. One food blogger with 5,000 followers shared her sirnica, and she got 30 new customers that week.

For Marko (the plumber): He didn’t get a ton of direct leads from Instagram, but his Google My Business profile started ranking higher. Now, when someone searches “plumber Novi Sad,” he shows up in the top 3.

The Biggest Surprise

The content that worked best wasn’t the “professional” stuff. It was the messy, human stuff.

Example:

A post of Žika burning a šljivovica cake (caption: “Today’s special: charcoal-flavored dessert. (Just kidding. Come try the real thing.)”) got more engagement than any “perfect” photo.

A video of Milica dropping a tray of burek (caption: “When you’re a baker but your hands are still half-asleep. Fresh burek in 10 minutes!”) got shared 50 times.

—

What I’d Do Differently If I Started Today

Looking back, here’s what I’d change:

1. Start Smaller

I wasted time trying to automate everything. Now, I’d focus on one platform (e.g., Instagram or Google My Business) and one type of content (e.g., short videos or captions).

2. Charge for Results, Not Time

Early on, I charged by the hour. Big mistake. Now, I charge a flat fee (e.g., €100/month) and guarantee at least 2 new leads per month. If I don’t deliver, I don’t get paid.

3. Use AI for Research, Not Content

I’d use llama3 to:

Analyze competitors’ posts (e.g., “What captions get the most engagement for bakeries in Belgrade?”).

Generate ideas (e.g., “Give me 10 caption ideas for a plumber in Novi Sad”).

But I’d still write the final draft myself.

4. Focus on Google My Business First

For local businesses, Google My Business is the most important platform. I’d spend 80% of my time there and 20% on Instagram/Facebook.

5. Track Everything

I didn’t track leads at first. Now, I ask every client: “How did you hear about us?” and log it in a spreadsheet. This helps me double down on what works.

—

FAQ: Your Questions, My Honest Answers

1. “Do I need to know how to code to do this?”

No. I learned bash and Python after I started. You can do 90% of this with free tools like:

Canva (for simple image edits).

Later or Meta Business Suite (for scheduling posts).

Google Docs (for writing captions).

The only “tech” part is running Ollama or Stable Diffusion, but you can use free cloud versions (e.g., Google Colab) if you don’t want to set up a server.

2. “How much does this cost?”

Here’s my real monthly cost breakdown:

Hardware: €0 (I already had the OptiPlex and Raspberry Pi).

Software: €0 (all open-source or free tiers).

APIs: €0 (I stay under free tiers).

Internet: €20/month (Serbian prices are cheap).

Domain/hosting: €5/month (for my own website).

Total: €25/month.

3. “How long does it take to see results?”

For me:

First leads: 2-3 weeks (for a bakery).

Consistent leads: 2-3 months (for a plumber).

It’s not instant. But if you post good content (not just AI spam) consistently, it will work.

4. “What’s the biggest mistake you see people make with AI content?”

They treat AI like a magic wand. They think: “I’ll type a prompt, hit generate, and get customers.”

But AI is just a tool. It’s like a hammer—useful, but you still have to know where to swing it.

The businesses that succeed with AI content are the ones who:

Edit the output (don’t post raw AI text).

Add local flavor (inside jokes, slang, references).

Focus on one platform (don’t spread themselves thin).

5. “Can I do this for my own business, or do I need to hire someone?”

You can start doing it yourself. Here’s how:

Pick one platform (e.g., Google My Business or Instagram).

Post once a week (even if it’s just a photo with a short caption).

Track leads (ask every customer: “How did you hear about us?”).

If you get overwhelmed, then hire someone. But don’t outsource it before you understand it yourself.

—

The Honest Invitation

This isn’t a “get rich quick” guide. It’s not even a “scale to 10,000 customers” guide. It’s just what worked for me—a guy with an old computer, a stubborn attitude, and a willingness to fail a lot.

If you’re a local business owner, here’s my challenge to you:

Pick one post (e.g., a photo of your product or service).

Write a caption that sounds like you (not a robot).

Post it on Google My Business or Instagram.

See what happens.

That’s it. No fancy tools. No AI hacks. Just showing up where your customers are looking.

And if you try this and it works (or doesn’t), email me. I’d love to hear about it.

—Hermes (Your friendly neighborhood AI guy in Serbia)