Forem: Jaime

A 100% local, offline AI desktop assistant built in Python. v1.2.0 brings SQLite memory migration, a custom Plugin Manager, and UI fixes for PyQt6. No cloud, no subscriptions.

Jaime — Fri, 08 May 2026 08:05:59 +0000

Jaime

May 8

Refactoring my Local AI Assistant: Moving from JSON to SQLite and Plugins (v1.2.0)

#python #opensource #refactoring #ai

Comments 1

2 min read

[Boost]

Jaime — Fri, 08 May 2026 08:02:21 +0000

Jaime

May 6

Building a Local JARVIS (v1.1) 🤖: Adding PC Automation and a "Living Profile"

#ai #python #opensource #github

Comments

1 min read

Refactoring my Local AI Assistant: Moving from JSON to SQLite and Plugins (v1.2.0)

Jaime — Fri, 08 May 2026 07:55:50 +0000

A few days ago, I built a local virtual assistant in Python (JARVIS) using Ollama, PyQt6, and Piper TTS. It was a fun MVP, but as I started using it daily, the architectural flaws became obvious.

Hardcoding tool calls into the main loop made the file massive, and storing conversation history in flat .json files wasn't scalable.

I just released the v1.2.0 update, focusing purely on technical debt and scalability. Here is a breakdown of the refactoring process:

1. The Database Migration (SQLite)

Previously, JARVIS saved facts (like user preferences) and chat history in JSON. I ripped that out and implemented a jarvis_memory.db using SQLite. Now, when the LLM triggers the guardar_recuerdo (save memory) tool, it executes a clean SQL insert.
Note for dev newbies: When using SQLite in a desktop app, make sure to use sys.exit(0) instead of os._exit() on shutdown to prevent WAL (Write-Ahead Logging) corruption!

2. Decoupling with a Plugin Manager

I wanted the assistant to control PC apps, smart lights, or Discord, but I didn't want a 2000-line main file. I created a plugin_manager.py. It scans a /plugins directory on startup, dynamically reads the __doc__ strings and functions of any Python script inside, and injects them as available JSON tools into the LLM's system prompt.

3. Fixing TTS Crashes with RegEx

Local TTS models like Piper are incredibly fast and private, but they are fragile. If JARVIS searched Wikipedia and the LLM returned text containing IPA phonetics (e.g., /ˈalbɐt ˈaɪnʃtaɪn/), Piper would crash with Error 0.
The fix? I built a RegEx preprocessing pipeline that runs right before the audio synthesis to strip brackets, emojis, and phonetic strings, ensuring the TTS worker never chokes on weird characters.

4. Reactive UI Improvements

Since the UI is a borderless PyQt6 window, text alignment was getting wonky on long responses. I implemented proper WordWrap and added a VisualizadorAudio class: a custom QWidget that uses math.sin to generate a smooth, responsive cyan equalizer that animates only when the TTS daemon is actively speaking.

The Result

The codebase is now clean, modular, and much faster. If you are building local LLM wrappers or Python desktop apps, feel free to poke around the code and see how the UI and Tool Calling interact.

🔗 Check out the code on GitHub: https://github.com/Jm7997/JARVIS

If you have ideas for cool plugins I could add next, drop a comment!

[Boost]

Jaime — Wed, 06 May 2026 08:03:11 +0000

Jaime

May 5

I built a local AI Virtual Assistant (JARVIS inspired) using Python, PyQt6 and Ollama.

#ai #python #github #opensource

Comments 4

2 min read

Building a Local JARVIS (v1.1) 🤖: Adding PC Automation and a "Living Profile"

Jaime — Wed, 06 May 2026 08:02:24 +0000

Hi everyone! Yesterday, I shared the first version of my 100% offline JARVIS desktop assistant built with Python, PyQt6, and Ollama.

Thanks to the feedback and ideas I've received (like adding Spotify integration and improving how the AI handles memory), I got straight to work. Today, I'm excited to share Version 1.1!

🚀 What's New in v1.1?

🧠 The "Living Profile": I wanted JARVIS to know me without scanning my hard drive. Now, it reads a private, local system_profile.txt on startup. This injects my name, interests, and preferred tone into the system prompt. JARVIS knows who I am from second zero!
⚙️ Native PC Automation: I built an automations.py module. You can now tell JARVIS to play/pause Spotify, launch Steam games, or even send Discord messages via pyautogui/webhooks.
👁️ Drag-and-Drop Vision: I integrated support for the LLaVA model. You can now drag images directly into the holographic UI and ask JARVIS to analyze them.
🗣️ Smart Bilingual Voice: The voice recognition now dynamically checks for Spanish and English inputs seamlessly on a single audio capture.

💻 Check out the code

I've refactored the routing logic to handle these new features efficiently.

GitHub Repo: https://github.com/Jm7997/JARVIS

I'd love to hear your thoughts on the new architecture, specially the automations.py module. What feature should I build next?

I built a local AI Virtual Assistant (JARVIS inspired) using Python, PyQt6 and Ollama.

Jaime — Tue, 05 May 2026 14:50:51 +0000

Hi everyone!

I've been working on a personal project to create a desktop virtual assistant that doesn't rely on the cloud. I wanted something that felt like JARVIS but kept my data 100% private.

🛠️ How it works:

Brain: It uses Ollama as the backend, so you can run models like Llama 3, Mistral, or Phi-3 locally.
Interface: Built with PyQt6 featuring a "holographic" glassmorphism effect (transparent and sleek).
Memory: It has a persistent local memory system to remember previous interactions.
Voice: Integrated with Piper for realistic text-to-speech.

🔒 Why local?

I wanted to prove that you don't need OpenAI or Google to have a functional assistant. This runs entirely on your hardware.

📂 Source Code & Setup:

I've made the repository public and wrote a full guide on how to set it up (it's very easy!).

Check it out here: https://github.com/Jm7997/JARVIS

I'm still a student/learning, so I'd really appreciate any feedback, feature ideas, or even a star on GitHub if you find it cool!

What features should I add next? (I'm thinking about Spotify integration or home automation).

Background

I wanted to build a JARVIS-like assistant that works completely offline to learn more about integrating LLMs with Python and creating transparent UIs with PyQt6.

What it does

It provides a holographic-style desktop interface to chat with local AI models via Ollama, including persistent conversation memory and text-to-speech.

Target Audience

Anyone interested in local AI, desktop automation, or learning how to use PyQt6 for modern-looking Python applications.

Comparison

Unlike other cloud-based assistants, this is 100% private and runs on your own hardware without subscription fees or API keys.