How a 75-Year-Old Retiree Built a Local AI (With a Face, Voice, and a Wiki Brain)

Building Your Own Offline AI: A Step-by-Step Guide by a 75-Year-Old Retiree

In the rapidly evolving world of artificial intelligence, it’s easy to assume that creating your own AI assistant is reserved for tech experts with years of coding experience. But what if I told you that, at 75 years old, with no programming background, I managed to build a fully functional, offline AI companion—with a face, voice, and even a knowledge wiki? And you can too.

Here’s my personal journey and the practical steps I took to bring this project to life, aiming to inspire others regardless of age or technical skills.

A Personal Disclaimer

Let me start by saying—I am not a coder. Python, Git, and AI models were all new to me. Before this project, “Git” was just something I said when I stubbed my toe. I enjoy growing weed, playing video games, and learning new things—nothing more.

Over the past week, I built a talking AI avatar using open-source tools that run entirely on my own computer, offline, with no subscriptions, APIs, or monthly fees. If someone like me can do it, so can you.

Understanding the Challenges and Preparation

Initially, my system was cluttered with background apps, crypto extensions, multiple browsers, and bloatware that bogged down my hardware. My GPU was idle while my CPU was overworked, leading to slow responses from the AI.

I spent time cleaning up my system:

Uninstalled unnecessary software (Dell bloatware, Alienware lighting, duplicate antivirus)
Closed Chrome and other resource-heavy programs
Disabled hardware acceleration and background processes

This cleared up my system’s resources considerably, enabling the AI to run more efficiently. I recommend doing the same: open Task Manager, identify resource hogs, and pause or close unneeded processes.

Hardware Context

For reference, I used robust hardware:

CPU: Intel Core i9-14900KF
RAM: 32 GB
GPU: NVIDIA GeForce RTX 4090 with 24 GB VRAM
Storage: 400 GB free space

While you don’t need such high specs, knowing what I used helps set expectations. Smaller models can run on less powerful machines—this guide is adaptable.

What You’ll Achieve

After following this process, you’ll have two AI systems working in tandem:

Both run locally on your machine, ensuring privacy and zero ongoing costs.

Part 1: Creating a Talking AI Companion (“Mao”)

Step 1: Prepare Your System

Before installing any software, you need to set up the tools that make everything possible:

Install Python
Download from python.org. Use versions 3.10–3.12 (avoid 3.13 due to known issues). During installation, check “Add Python to PATH”. Verify installation by opening Command Prompt and typing python --version.
Install Git
Download from git-scm.com. Verify with git --version.
Install FFmpeg (for voice output)
Download from gyan.dev. Extract to C:\ffmpeg. Add C:\ffmpeg\bin to your system PATH via Environment Variables. Confirm with ffmpeg -version in a new Command Prompt.

Step 2: Install LM Studio and Download a Model

Visit lmstudio.ai and install the latest version compatible with your OS.
Download a suitable AI model from Hugging Face: bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF
(or a smaller model if preferred). Place it in LM Studio’s models folder.

Step 3: Configure LM Studio

Launch LM Studio, select your model.
Set GPU Offload to maximum for faster responses.
Adjust Context Length to 4096 for speed.
Enable Limit model offload to dedicated GPU memory.
Load the model and start the local inference server.

Step 4: Set Up the Voice and Face Avatar

Clone the Open-LLM-VTuber repository via Command Prompt:

bash git clone https://github.com/Open-LLM-VTuber/Open-LLM-VTuber cd Open-LLM-VTuber python -m venv venv venv\Scripts\activate pip install uv uv sync git submodule update --init --recursive copy config_templates\conf.default.yaml conf.yaml

Edit conf.yaml with Notepad, setting the model’s local URL and voice options:

yaml llm_provider: "ollama_llm" ollama_llm: base_url: "http://localhost:1234/v1" model: "deepseek-r1-distill-qwen-14b" tts_model: "edge_tts"

Launch the server:

bash uv run run_server.py

Open your browser at http://localhost:12393. You should see your avatar, ready to chat.

Troubleshooting

If you encounter connection errors (e.g., WebSocket issues), you may need to set a specific WebSocket URL in the browser console:

Press F12, go to Console, and run:

javascript localStorage.setItem('wsUrl', 'ws://127.0.0.1:12393/client-ws')

Refresh the page, and your avatar should reconnect.

Part 2: Building a Document and Wiki Answer Bot (“The Wiki Bot”)

Step 1: Install Ollama and Other Tools

Download and install Ollama.
Open Command Prompt and pull a smaller DeepSeek model:

bash ollama pull deepseek-r1:7b

Download and install AnythingLLM, a native desktop app that manages your knowledge base.

Step 2: Set Up a Workspace

Launch AnythingLLM, create a new workspace (e.g., “My Wiki”).
In settings, choose Ollama as your provider, select the deepseek-r1:7b model.

Step 3: Connect the Browser Extension

Install the AnythingLLM browser extension from Chrome Web Store.
In Desktop settings, generate an API key, which gives you a connection string like:

http://<your_api_key>@localhost:3001

Enter this full connection string into the extension’s configuration.

Step 4: Populate Your Knowledge Base

Visit your favorite wikis, PDFs, or websites.
Use the extension icon to “Send entire webpage” or upload files directly into the workspace.
The bot now has access to your documents, ready to answer questions contextually.

Step 5: Ask Your Bot

In AnythingLLM, type questions. It will answer based solely on the content you’ve provided, citing sources.

Handy Tips and Common Fixes

Cost and Time Investment

This entire setup is free—using open-source tools and your existing hardware.
Initial setup may take between 15-40 hours, depending on experience. I spent about 20 hours building and troubleshooting.
Once established, your AI systems can run continuously without ongoing fees.

Final Words of Encouragement

You don’t need a computer science degree or expensive cloud subscriptions. Patience, curiosity, and a willingness to troubleshoot are your best assets. If I, at 75, can do this, anyone can—regardless of age or background.

I hope this guide proves helpful. Feel free to reach out with questions—I love hearing about others building their own AI companions. Happy building!

— Huanchaquero

How a 75-Year-Old Retiree Built a Local AI (With a Face, Voice, and a Wiki Brain) — And You Can Too

Part 1: Creating a Talking AI Companion (“Mao”)

Part 2: Building a Document and Wiki Answer Bot (“The Wiki Bot”)

Handy Tips and Common Fixes

Cost and Time Investment

Final Words of Encouragement

Comments

Leave a Reply Cancel reply