Browse 37 articles on AI engineering, machine learning, LLMs, MLOps, and modern software development. Written by Misha Lubich, AI Engineer & Technical Leader with experience at Apple, GitHub, and cutting-edge AI startups.
A field report from a quarter where the demos looked great, the dashboards looked calm, and the agent stack quietly set small piles of money on fire.
Category: AI Architecture | Reading time: 7 min | Tags: Agents, Architecture, Reliability, Postmortem, LLMOps
By Misha Lubich
AI UX gets weird fast when every feature invents its own loading state, confidence label, and apology paragraph. A design system keeps the product trustworthy.
Category: AI Products | Reading time: 3 min | Tags: UX, Design Systems, AI Products, Interaction, Frontend
By Misha Lubich
A renamed field should not take down an enterprise agent rollout. MCP only scales if contracts are tested like APIs, not treated like vibes with JSON.
Category: Open Source | Reading time: 3 min | Tags: MCP, Contracts, Testing, Enterprise, Integration
By Misha Lubich
June's AI market feels less like a breakthrough month and more like the morning after a very expensive demo party. The winners are building boring control systems.
Category: Hot Takes | Reading time: 4 min | Tags: AI Industry, Agents, Strategy, Startups, LLMOps
By Misha Lubich
A practical runbook from debugging an agent stack where HTTP was green, dashboards were calm, and the agent was quietly doing interpretive dance with tool calls.
Category: MLOps | Reading time: 4 min | Tags: Observability, Agents, Reliability, Runbook, Monitoring
By Misha Lubich
A RAG answer can be correct and still lose the user. I learned that the boring way: by watching a good retrieval pipeline feel slow enough to be broken.
Category: AI Products | Reading time: 3 min | Tags: RAG, Latency, Performance, Architecture, Product
By Misha Lubich
If your AI budget has tokens but no eval line item, you did not make a budget. You made a very confident wish with a model invoice attached.
Category: MLOps | Reading time: 4 min | Tags: Evals, Budgeting, Quality, Governance, LLMOps
By Misha Lubich
Teams over-invest in model upgrades while stale retrieval quietly destroys answer quality. Fresh evidence often beats a larger checkpoint.
Category: AI Architecture | Reading time: 1 min | Tags: RAG, Retrieval, LLM, Data Freshness, Search
By Misha Lubich
Most AI postmortems read like blame theater. A useful one produces guardrails, eval cases, and a measurable drop in repeat incidents.
Category: Engineering Culture | Reading time: 1 min | Tags: Postmortems, Agents, Reliability, Process, Leadership
By Misha Lubich
Long-context models tempt teams to treat the prompt as a database. That works until you need auditable state, incremental updates, and retrieval that survives a page refresh.
Category: AI Architecture | Reading time: 2 min | Tags: Context Windows, RAG, State, LLM, Architecture
By Misha Lubich
Demos optimize for applause. Production optimizes for margin, rollback, and an angry user with a spreadsheet. Here is the checklist I use before calling something shipped.
Category: AI Products | Reading time: 1 min | Tags: Product, Launch, LLMOps, Governance, Checklist
By Misha Lubich
The model says the row was updated. The audit log disagrees. Until you treat tool I/O like distributed systems, agents will keep shipping confident lies.
Category: MLOps | Reading time: 1 min | Tags: Agents, Tool Calling, Reliability, Observability, Production
By Misha Lubich
Most AI products don't die from lack of demos; they die from unit economics nobody modeled. Cost discipline is now a core architecture decision.
Category: AI Products | Reading time: 2 min | Tags: AI Costs, Unit Economics, LLMOps, Product Strategy, Architecture
By Misha Lubich
OpenClaw turned personal AI agents from a demo into something people actually run: a local-or-near-local control plane wired into the apps you already live in. Here is how it works, where it shines, and where it can burn you—illustrated with architecture diagrams and honest tradeoffs.
Category: Open Source | Reading time: 9 min | Tags: OpenClaw, AI Agents, Messaging, Security, Self-Hosting
By Misha Lubich
The fastest way to lose trust in an AI feature is shipping it with vibes and no evaluation harness. In 2026, release quality is mostly decided before launch day.
Category: AI Architecture | Reading time: 2 min | Tags: AI Evaluation, LLM, Quality, Testing, Production
By Misha Lubich
I wired up my first MCP server on a Sunday. By Tuesday I believed I'd solved tool calling forever. A month later I was drawing boxes on a whiteboard about auth, gateways, and who exactly gets sued if the agent deletes the wrong row.
Category: AI Architecture | Reading time: 4 min | Tags: MCP, API Design, Security, Agents, Production
By Misha Lubich
Everyone's talking about agentic coding in 2026. The charts look great. But if you actually ask engineers what they're willing to hand off end-to-end, the room gets quiet. That gap isn't hypocrisy — it's the whole story.
Category: Engineering Culture | Reading time: 5 min | Tags: Agentic AI, Claude, Engineering Leadership, AI Tools, Process
By Misha Lubich
Andrej Karpathy coined 'vibe coding' and Twitter loved it. But building production systems by vibes is how you get production incidents by vibes.
Category: Hot Takes | Reading time: 3 min | Tags: Vibe Coding, AI Coding, Best Practices, Hot Takes, Karpathy
By Misha Lubich
Hot takes declared RAG dead. Long-context models were supposed to replace it. But in early 2026, Cursor is shipping RAG pipelines, engineers are still optimizing chunking, and retrieval is evolving — not dying. Here's what's actually happening.
Category: AI Architecture | Reading time: 6 min | Tags: RAG, Long Context, LLM, Architecture
By Misha Lubich
After 6 months of using Cursor as my primary IDE, my velocity has tripled. Agentic workflows—not comment-driven inline gen—are what actually moved the needle.
Category: AI Products | Reading time: 3 min | Tags: Cursor, AI Coding, Developer Tools, IDE, Productivity
By Misha Lubich
LangChain was the jQuery of AI — necessary for a moment, then a liability. Modern AI engineering demands less abstraction, not more.
Category: AI Architecture | Reading time: 7 min | Tags: LangChain, Frameworks, Simplicity, Architecture
By Misha Lubich
OpenAI's o3 and DeepSeek's R1 proved that chain-of-thought at inference time is the next frontier. Here's what this means for how we build AI systems.
Category: AI Architecture | Reading time: 2 min | Tags: o3, DeepSeek, Reasoning, Chain of Thought, Architecture
By Misha Lubich
CrewAI, AutoGen, and LangGraph promise autonomous agent teams. I deployed all three to production. Here's the unvarnished truth about what works and what's pure marketing.
Category: AI Architecture | Reading time: 3 min | Tags: CrewAI, Multi-Agent, AutoGen, LangGraph, Production
By Misha Lubich
We're building the last generation of hand-written CRUD apps. AI agents will replace 80% of backend code within 3 years. Plan accordingly.
Category: AI Architecture | Reading time: 7 min | Tags: AI Agents, Architecture, Future, Software Engineering
By Misha Lubich
I've tried every AI coding CLI — Aider, Mentat, GPT-Engineer. Claude Code is the first one I trust to make changes across a real codebase without supervision.
Category: AI Products | Reading time: 3 min | Tags: Claude Code, Anthropic, Terminal, AI Coding, Developer Tools
By Misha Lubich
The industry created a fake job title to make 'writing instructions for a chatbot' sound like a technical discipline. Let's stop pretending.
Category: Hot Takes | Reading time: 2 min | Tags: Prompt Engineering, Career, Hot Takes, Industry
By Misha Lubich
The entry-level programming job as we know it is disappearing. This is the most important conversation our industry refuses to have.
Category: Engineering Culture | Reading time: 8 min | Tags: AI Coding, Junior Developers, Career, Future of Work
By Misha Lubich
We switched from GPT-4o to fine-tuned Llama and cut our costs by 94%. Our quality scores went up. Here's the playbook.
Category: Open Source | Reading time: 2 min | Tags: Open Source, Llama, Cost Optimization, Self-Hosting
By Misha Lubich
Companies are spending millions on AI teams that produce impressive demos and zero production value. I've seen it from the inside.
Category: Engineering Culture | Reading time: 8 min | Tags: Hiring, AI Teams, Management, Productivity
By Misha Lubich
We've gotten incredibly good at building AI systems. We're still terrible at knowing whether they actually work. Evals are the bottleneck nobody's fixing.
Category: MLOps | Reading time: 3 min | Tags: Evaluation, Testing, Quality, MLOps, Best Practices
By Misha Lubich
That fancy Kubeflow/Airflow/Prefect ML pipeline you built? It's the most expensive, fragile, and unnecessary code in your entire stack.
Category: MLOps | Reading time: 2 min | Tags: MLOps, Technical Debt, Pipelines, Infrastructure
By Misha Lubich
Anthropic's Model Context Protocol is the USB-C of AI tooling. Once adoption hits critical mass, every custom integration layer becomes unnecessary.
Category: AI Architecture | Reading time: 2 min | Tags: MCP, Protocols, Anthropic, Standards
By Misha Lubich
The industry cargo-culted microservice architecture into ML platforms and created distributed systems nightmares. Monoliths are the answer.
Category: MLOps | Reading time: 2 min | Tags: Microservices, Architecture, ML Systems, Infrastructure
By Misha Lubich
Most AI safety work is performative theater designed to look responsible while not actually slowing anything down. Let's have an honest conversation.
Category: Hot Takes | Reading time: 2 min | Tags: AI Safety, Ethics, Industry, Regulation
By Misha Lubich
90% of 'AI-powered' startups are solutions searching for problems. The graveyard of AI products is full of technically brilliant ideas that nobody needed.
Category: AI Products | Reading time: 2 min | Tags: Product, Startups, Strategy, Market Fit
By Misha Lubich
Every company will need fine-tuned models within 18 months. The problem is that 95% of fine-tuning efforts fail because teams treat it like training from scratch.
Category: AI Architecture | Reading time: 2 min | Tags: Fine-Tuning, LLM, Training, Best Practices
By Misha Lubich
Teams waiting for the next model release to fix their broken AI products are deluding themselves. Your architecture is the bottleneck, not the model.
Category: AI Architecture | Reading time: 2 min | Tags: Architecture, LLM, System Design, Best Practices
By Misha Lubich
Controversial takes, hard-won lessons, and unfiltered opinions on modern AI engineering. No hype — just what actually works in production.
