Grevire — from the Fribourg patois for “Gruyère” — is a name that reflects my roots in the Gruyère region, in the canton of Fribourg, Switzerland. My mission? Documenting my personal R&D on local AI, autonomous agents, and sharing real-world discoveries without commercial bullshit.
Who am I?
Mathieu, aka GrevireDev. Wearing two hats for 17 years: paramedic consultant specializing in pre-hospital care by day, freelance developer under the name Grevire by night.
My dev journey? Classic yet chaotic. Discovered code at 17 with the HTML/CSS/JS and PHP trio, then Python, Node.js, and now anchored in the SvelteKit + Node.js + Electron ecosystem. Linux environment (Arch-based), Proxmox for the HomeLab, VPS cluster for projects that demand more power.
This blog was born from a simple observation: in the French-speaking AI world, we often sell THE miracle solution at premium prices, with closed approaches that don’t match the open source philosophy. There’s a crucial lack of objective, field-tested feedback with real comparisons between models and tools.
My philosophy? “I test, I document, I share” — learning through hands-on projects, not YouTube tutorials.
What you’ll find here
1. Mini PC tests and benchmarks for local AI
Functional benchmarks based on reproduced standard actions. No synthetic scores like 3DMark, but concrete stuff: how long to process a given prompt, what response quality, what system stability.
2. Autonomous agents documentation and AI R&D
Focus on real user experience. How does it actually work when you move beyond proof-of-concept? Concrete example: my search for a messaging gateway for Hermes Agent (Telegram → Discord → Mattermost → Hermes WebUI) — nobody documented these comparisons in a structured way.
3. Development experience feedback
Tools, workflows, problems solved… and those that failed. Because failure is part of the process and deserves to be documented as much as success.
4. Comparative AI model analysis
Local and cloud-based. Kimi 2.6, MiniMax 2.5, Mistral via OpenRouter… I test, compare, and document real usage differences.
Frequency: 2-3 posts per week Workflow: Voice dictation → AI-assisted writing → automated publishing (yes, I use AI to document AI)
My testing approach
The hardware
GMKtec K8 Plus: AMD Ryzen 7 8845HS, 8c/16t, 5.1 GHz, DDR5 32 GB, PCIe 4.0 1TB SSD, iGPU Radeon 780M. No dedicated GPU — everything runs on the iGPU with relatively low inference rates on Ollama, but that’s intentional.
The methodology
Functional benchmarks reproducing real business tasks. No question of measuring FLOPS or tokens/second in a vacuum. I reproduce real scenarios: article writing, code analysis, assisted debugging, image generation for the blog.
Community & exchanges
“I document to share, I share to learn”
This blog is aimed at junior developers seeking inspiration, seniors looking for concrete use case analysis, and anyone interested in local AI without wanting to get scammed by overpriced proprietary solutions.
My approach isn’t that of an infallible expert who spouts truths. It’s that of a practitioner who experiments, documents discoveries and mistakes, and encourages mutual learning.
Challenge my results. Propose your own tests. Share your failures as much as your successes. That’s how we collectively advance in this constantly evolving ecosystem.
Contact me
- LinkedIn: Mathieu Brocas
- X/Twitter: @GrevireDev
- GitHub: brocasm