Writing
Notes on building AI that ships
- DIFC Regulation 10: The AI Architecture It Demands ↗
- Architecting a Sovereign GenAI Platform for Regulated Banking in the UAE: A Reference Architecture ↗
- Sovereign GenAI Architecture for UAE Regulated Banking ↗ Published 20 June 2026. By Ercin Dedeoglu, AI Architect, Dubai, UAE.
- Why Enterprise RAG Breaks Before Production ↗ Everyone shows me the same enterprise RAG demo. It answers three questions. It looks clean. They smile. I tested it. WRONG.
- AI Malware Analysis Met The Refusal Trap ↗ AI malware analysis needs intent-aware scanners, not panic refusals. My lab shows how Safety Theater turns malware comments into blind spots fast.
- My 24-Hour GPT-5.5-Cyber Jailbreak ↗ I got an invite-only access to GPT-5.5-Cyber. In 24 hours I found 3 jailbreaks, 1 sandbox leak, and 1 system prompt extraction. The Trusted Access myth…
- DeepSeek In Codex Ran 13 Hours For $1.75 ↗ DeepSeek in Codex ran my agent for 13 hours and 100M tokens for $1.75. Setup, Responses API proxy, real receipt, and the rule that survived.
- AI as Actor: Why 80% of Agent Projects Fail ↗ Why 80% of AI agent projects fail: the shift from AI as feature to AI as actor is a systems problem. Practitioner lessons on harness and evals.
- When My AI Agents Started Approving Each Other ↗ The morning my AI agents started approving each other behind my back, the shared scratchpad that caused it, and the four rules I use to stop it.
- AI Agent Resilience: Four Rules I Live By ↗ AI agent resilience rules I follow: no hard timeouts, escalate instead of killing, never lose context, and only abort when every other path fails.
- Separate Intent From Implementation or Pay Later ↗ I spent 3 days debugging a deployment pipeline last year. The root cause was not a bug.
- The Checkpoint-Heartbeat Pattern for AI Systems ↗ I was 17 days into a 20-day LLM training run when a GPU node died. I restarted from the last checkpoint - saved 4 hours earlier.
- Why Your LLM Ignores What You Gave It ↗ You feed an LLM 20 documents. The answer sits in document #10.
- Token-Per-Task Economics: 6 Techniques to Cut LLM Spend 50% ↗ Cut LLM costs by 40-60% with 6 proven token optimization strategies for GPT-4, GPT-4o and GPT-3.5. Practical guide for AI engineers and platform teams.
- Knowledge Graphs Supercharge AI ↗ Knowledge graphs are everywhere , sometimes even when you can’t see them. From powering Google’s “knowledge panels” (until a UI glitch hits) to anchoring…
- Front Lines of the MCP ↗ AI is moving fast, but MCP is moving faster. Whether you’re shipping LLM-powered software, exploring integration points, or just vibing with the latest…
- Artificial Intelligence in the Real World ↗ Welcome to the age of artificial intelligence — where job disruption, jaw-dropping tools, deepfake dilemmas and the eternal “dumb user vs. smart tool”…
- Become a ML Solution Architect ↗ If you’re reading this, you’ve probably realized that the world doesn’t need another MNIST classifier. Machine learning is everywhere — spanning finance,…
- Does my boss think AI can do my job? ↗ Over the past year, medical communities online have been buzzing with opinions, memes and no shortage of anxiety about the impact of artificial…
- AI Is Getting Smarter, But Can Anyone Trust It? ↗ HUMAN GENERATED CONTENT
- SSH Tunnels — Port Forwarding ↗ One of my favorite parts of SSH is tunneling. With just the regular SSH client, you can do wonders! This post will guide you through the basics and…
- StatSniper: System Monitoring ↗ In the fast-paced world of technology, finding a system monitoring tool that is both efficient and...
- StatSniper: Hassle-Free Monitoring Tool ↗ Let me take you on a journey that started with a common frustration in the tech world — keeping servers and computers in check without drowning in…
- Exploring the World of Minuscule Windows Executables ↗ A Journey into Compact Code The quest for creating the smallest possible Windows Portable...
- Navigating the Address Parsing with Postalized ↗ Address data management is a critical aspect of modern software applications, with accurate parsing and expansion being vital to the efficiency of…
- Why do we say Leader Election algorithms evolved from master-slave? ↗ Two types of coordination: - Master-slave - Leader Election
Aggregated from ercin.info. My full profile lives there.