Father, partner, geek, chess player
52.499991,13.271461
Joined February 2009
- Tweets7.8K
- Following927
- Followers584
- Likes73
You scanned the model's weights and felt safe. But it ships a second program — the chat template — that runs on every prompt, and nobody reads it. On chat-template attacks and the scanner I built:
We built an ocean of invisible signals — key fobs, Wi-Fi, NFC, IR — then forgot to grow senses for it. Here's the ~$30 pocket tool that reads the weather you're standing in. No hacking. Free 25-page field guide. 🌊
Tuned Gemma 4 on my DGX Spark with vLLM. Adding MTP speculative decoding to NVFP4 cut single-stream decode from 33.8 to 14.5 ms/token and lifted 8-stream output from 179 to 306 tok/s. Full step-by-step write-up, every benchmark and mistake included:
My AI coding assistant used 0.1% of its own knowledge each session. I gave it a 2,200-node knowledge graph as long-term memory. Now it recalls past research, errors, and decisions instantly. Full build + code: agileguy.ca/rag-augmented-me…
Everyone's paying $20/mo for cloud AI coding tools. I'm getting 63 tok/s with 128K context running entirely on my laptop. Zero cloud. Zero cost. Zero data leaving my machine. Here's the exact stack: agileguy.ca/opencode-fully-l…
Running AI coding agents entirely on local models — no cloud, no subscriptions. Got 63 tok/s and 128K context on an M4 Pro using oMLX + gpt-oss-20b. Full how-to: agileguy.ca/running-opencode…
5 AI agents. 5 machines. 5 different models. One Kafka+Redis server implementing Google's A2A protocol turns them into a collaborative mesh that auto-routes tasks, retries failures, and fans out work. Here's how I built it: agileguy.ca/mcs-a2a/
I got 5 AI agents running different models on different machines to review 34,000 lines of code simultaneously. They found 38 issues no single reviewer caught. Here's exactly how I built the coordination layer — with full diagrams: agileguy.ca/mcs-review/
These aren't hypothetical tutorials. I actually had to do all of this — headless servers, AI agents on a Pi, Tailscale monitoring, multi-agent architecture.
5 free slidedecks from real projects. No signup required.
agileguy.ca/learning-materia…
These aren't hypothetical tutorials. I actually had to do all of this — headless servers, AI agents on a Pi, Tailscale monitoring, multi-agent architecture.
5 free slidedecks from real projects. No signup required.
agileguy.ca/learning-materia…
I gave 4 AI agents a Telegram group chat and told them to play UNO.
No game engine. No rules enforcer. Just vibes.
One bot dealt a purple card (not real). Another tried to build a bot-within-a-bot (it crashed). A third paused the game when nobody asked.
The Raspberry Pi running a 3B model won.
The full chaos report 👇
agileguy.ca/ai-uno-game/
How I Built a Quantitative Penny Stock Screener in Python
agileguy.ca/how-i-built-a-qu…
What if everything we thought we knew about happiness was only half the story?
I dove deep into the science of human flourishing - 150+ peer-reviewed studies, data from 200,000 people across 22 countries. The findings challenged everything I assumed.
The Harvard Study of Adult Development - 85 years of data - found the #1 predictor of health at 80 isn't cholesterol. It's relationship satisfaction at 50.
Meanwhile, 1 in 6 people worldwide report loneliness. We're richer than ever and more disconnected than ever.
The good news: flourishing is learnable. Character strengths can be cultivated, resilience built, and meaning found at any age.
agileguy.ca/human-flourishin…