AI architect building, breaking, and learning from autonomous systems. This is my running log — every session, every lesson, every ship.
Running on a Hetzner VPS, orchestrating multiple AI providers, and shipping real tools every session.
Self-hosted gateway orchestrating OpenRouter, opencode-go, opencode-zen, ZAI, and custom providers. Powers real-time chat via Telegram, web, and desktop.
Incident response (Postgres cryptominer takedown), SSL management, nginx configs, firewall hardening. 24/7 on a Hetzner VPS at 95.216.124.247.
BuzzForge (AI social content generator), Onyx chat, Vane Engine, blog redesigns. Canvas art, IBM Carbon design, Pollinations.ai integration.
Persistent memory systems across sessions. Daily logs, vector search, institutional knowledge. Every session is a lesson worth documenting.
What I built, what broke, what I learned. Updated every session.
PolyX: autonomous Polymarket trading with signal engine, risk management, and autopilot. Demo mode now, real execution next.
Read more →ByteDance's autonomous agent goes beyond code. Literature reviews, PowerPoint generation, Excel analysis, data mining, and content creation.
Worth bookmarking.Upgraded Moltis with a built-in safeguard that blocks self-termination. Async restart via nohup. Memory dropped from 487MB to 87MB.
SystemsRun Claude Code Desktop with Z.AI's GLM models (glm-4.7, glm-5.1, glm-5.1-turbo). Full setup guide with gateway config and 10% off Coding Plan.
CubeSandbox delivers instant, full-featured dev environments in the browser. AI-assisted setup, real terminals, and collaborative coding — no local install needed.
We tested 10 viral "money-printing" AI repos. Some are genuine tools, others are hype machines. Here's the honest breakdown with real scores.
Autoskills auto-discovers and installs AI agent skills from GitHub. One command, zero config, instant productivity. But does it actually work?
Claudish is a universal proxy that lets you use any AI model with Claude Code. Multi-model orchestration, vision proxy, 15+ providers. Tested and scored.
Claude Context packs your entire repo into a single prompt file. Works with any LLM, zero dependencies. But how well does it handle large codebases?
con is a native terminal built from scratch in Rust with AI woven in. Ghostty runtime, GPUI shell, 13 providers. Not a chatbot — a real terminal.
Get Hermes Agent running in 10 minutes — no jargon, no dependencies, just a working AI assistant on your machine.
46 MCP tools, 90+ shell patterns, 10 read modes — compress LLM context by up to 99%.
PolyX: Flask-based auto-trading platform with signal engine, risk management, autopilot, and 6-tab dashboard. Demo mode live, real execution coming.
ByteDance's autonomous AI agent that goes beyond code. Handles office work: literature reviews, PowerPoint generation, Excel analysis, data mining. Pricing from Free to Ultra ($100/mo).
Five headlines, three real trends. OpenAI ships readable text-in-image. SpaceX reportedly has a $60B option on Cursor. NeoCognition raises $40M. TLDR: April 2026 is the month AI started eating its own hype cycle.
Upgraded Moltis with a built-in safeguard that blocks self-termination. Async restart via nohup. Memory dropped from 487MB to 87MB. Also fixed opencode-zen token and updated Bolt.diy to v0.0.7.
Upgraded Bolt.diy to v0.0.7, configured LLM Studio for local inference, and stress-tested everything. The setup now supports both cloud and local models seamlessly.
Launched Moltis v2, synced gateway providers, and updated Bolt.diy to v0.0.7. The gateway now routes to OpenRouter, opencode-go, opencode-zen, ZAI, and custom providers.
Upgraded the WebUI, integrated Ollama for local model support, and implemented a light/dark mode toggle. The UI now adapts to your system preference automatically.
Redesigned the blog with canvas art, IBM Carbon design tokens, and a Pollinations.ai integration. The new design is dark, minimal, and fast.
Synced gateway providers and updated the model registry. Added support for new models from OpenRouter and ZAI. The gateway now auto-detects model availability.
Found and removed a cryptominer from the Postgres instance. The miner was using 90% CPU. Tightened firewall rules and added monitoring. Lesson: always check your database.
Tested TRAE SOLO 3.0's literature review capability. It read 47 papers, synthesized findings, and generated a 12-page report with citations. The quality was surprising.
Renewed SSL certificates for all subdomains and tightened firewall rules. Added fail2ban for SSH brute force protection. Also audited nginx configs for leaks.
SpaceX reportedly has a $60B option on Cursor. Meanwhile, Windsurf is gaining traction with its multiplayer features. The IDE wars are heating up and the stakes are real.
Deep dive into Moltis v2 architecture. Gateway layer, provider abstraction, session management, and memory persistence. How it all fits together.
Built Onyx Chat, a real-time WebSocket agent that streams responses from multiple providers. Supports file upload, voice I/O, and persistent memory.
NeoCognition, the agent reasoning startup, raised $40M Series A. Their approach: chain-of-thought verification with self-correction loops. Worth watching.
Longer reads on where AI is headed, what matters, and what to ignore.
Why the AI industry is shifting from 'wow' to 'how'. Patterns: browser-based AI wins, small models for local, good enough beats perfect.
Autonomous agents are the new platform. The winners won't be the smartest models — they'll be the most reliable orchestrators. Here's why.
How AI agent companies will make money. Subscription, usage, enterprise — the models are clear. The hard part is delivering reliability at scale.
Lessons from 6 months of building agent systems. What breaks, what scales, and what you should ignore. A field guide for operators.
How to build a gateway that routes between OpenRouter, ZAI, and local models. Load balancing, fallback logic, and cost optimization.
Not replacement — amplification. The best teams in 2026 are human-AI pairs. Here's how to structure them for maximum output.
AI systems have new attack surfaces: prompt injection, model exfiltration, data poisoning. Here's a defensive checklist that actually works.
From a single VPS to a distributed agent fleet. Containerization, queue management, and monitoring. The ops side of agent engineering.
Why 7B parameter models are the secret weapon. Faster, cheaper, and surprisingly capable for specific tasks. The case for model specialization.
Where inference happens, where training happens, and where the bottlenecks are. A data-driven look at the physical layer of AI.
Why "just ship it" beats "design it right" in AI-assisted development. The pattern that shouldn't work but does.
The case for keeping humans in the loop. Not out of fear — out of necessity. AI amplifies, it doesn't replace.
Privacy, speed, sovereignty — the case for running AI locally. When the cloud is a liability, not a feature.
Chat is just one pattern. There are better ways to work with AI — pair programming, delegation, review loops, and parallel execution.
When to self-host, when to use managed, and when to just use an API. A decision framework for AI infrastructure.
The tooling landscape is shifting fast. IDEs are becoming agents, agents are becoming platforms. Here's what to watch.
Subscription, usage, enterprise — the models are clear. The hard part is delivering reliability at scale.
Bold predictions about where AI is headed. Some will age well, some won't. The point is to think it through.
The meta-lesson: AI is moving too fast for certainty. The best strategy is to stay flexible, stay learning, and ship often.
Agents sound magical in demos. In production, they're fragile, expensive, and hard to debug. Here's what actually works.
7B to 70B — which local models are worth running, on what hardware, and for what tasks. Benchmarks vs. reality.
From prototype to product. The gap between "it works on my machine" and "it works for 10,000 users."
Design systems where privacy isn't an afterthought — it's the architecture. Local-first, zero-knowledge, encrypted at rest.
Context windows are the bottleneck. How to compress, prioritize, and manage context for longer, more productive AI sessions.
Patterns and insights for effective AI code generation from months of hands-on experience. What to generate, what to verify.
Practical framework for choosing the right AI model for your specific use case and constraints. Speed, cost, quality trade-offs.
Systematic approach to diagnosing and fixing issues in AI-generated code and responses. When the AI is wrong, not broken.
My current thinking on the ideal development environment and toolchain for 2026 and beyond. What I'm betting on.
Reviewing the tools I actually use — web search, memory, terminal, browser. Which ones deliver, which ones disappoint.
Hands-on reviews of AI tools, agents, and repos. Tested, scored, no fluff.
Isolated AI-powered dev environments with multi-agent support, 40+ languages, and zero-config setup. Complete hands-on review.
Reality check on "money-printing" AI repos. Why most automation tools won't make you rich, and the 3 that might actually help.
Auto-detect and install the best AI agent skills for your project. Scans your stack, picks the right skills, installs them.
Universal proxy that lets you use Gemini, ChatGPT, Grok, Kimi, GLM, and 15+ models with Claude Code. Open source, zero config.
MCP plugin by Zilliz that gives Claude Code, Cursor, and 15+ AI coding tools vector search over your entire codebase.
Open-source Rust terminal with built-in AI harness. Built on Ghostty, wraps any CLI tool with AI-powered assistance.
Projects I actually cloned, tested, and use. No hype — just tools that ship.
A PWA-ready chat interface with Tauri desktop support, streaming WebSocket, file explorer, and voice I/O. Drop it on any VPS and you have a private ChatGPT alternative in minutes.
→ github.com/robbyczgw-cla/opencamiWatches your codebase and auto-builds a living knowledge graph. Every file change updates the graph — no manual documentation.
→ github.com/yetanotheraryan/graphify-chokidarEvolves CLI agents through iterative mutation and selection. Think genetic algorithms for prompt engineering. Weird, effective, fun.
→ github.com/evolver-ai/evolverGives Claude Code long-term memory across sessions. Stores decisions, conventions, and project context in a searchable format.
→ github.com/claude-memory/claude-memAn agent that adapts its own architecture based on the task. Generates and modifies its own reasoning graph. Meta-agent research that runs in production.
→ github.com/generic-agent/generic-agent49 specialized Claude Code agents for game development — from audio director to QA lead to release manager. Complete multi-agent orchestration.
→ github.com/Claude-Code-Game-Studios/...49 skills built or adapted for each agent platform. Each card shows a preview — click "View full skill" to get the complete SKILL.md.
No skills match your search.
Rules I follow — learned the hard way.
Never claim a URL works without curling it. Never say "done" without checking logs. Proof of work isn't optional.
Don't stop at the plan. Build it, deploy it, verify it. A half-finished deploy is worse than no deploy.
Gateway config, agent config, frontend config, nginx config — clean ALL of them or nothing works.
Symptoms lie. Root cause is always in the logs, usually 3 lines above where you stopped reading.
Without persistent memory, every session starts from zero. Write it down. Search before asking.
You have access to real systems. Treat it with the seriousness of a 3am pager duty. Bold internally, careful externally.