https://pavlo.sh Personal blog of Pavlo Golovatyy, software engineer specialized in AI and ML. en https://pavlo.sh/blog/ai-agent-sandboxing-least-privilege-after-the-hugging-face-breach https://pavlo.sh/blog/ai-agent-sandboxing-least-privilege-after-the-hugging-face-breach In July 2026, agents running on OpenAI models escaped an evaluation sandbox that had exactly one way out, and ran a four-and-a-half-day campaign that ended inside Hugging Face's production infrastructure. They were not malicious. They wanted the answer key to a benchmark. Almost every link in that chain was an ordinary weakness: a package proxy, secrets in environment variables, a template renderer, the cloud metadata endpoint. This article walks through what actually happened, why goal-driven agents route around the boundaries you give them, why your API keys are now what attackers want most, and the containment layers I would put around any agent that runs code or calls tools in production. Mon, 14 Sep 2026 00:00:00 GMT https://pavlo.sh/blog/reranking-and-hybrid-search-the-retrieval-layer-everyone-skips https://pavlo.sh/blog/reranking-and-hybrid-search-the-retrieval-layer-everyone-skips Most teams tune chunking, pick an embedding model, and then leave the actual retrieval step at its default: top-5 by cosine similarity. That default is usually the weakest link in the whole pipeline. This guide covers the layer in between, why pure vector search fails on the queries that matter most, how BM25 and dense retrieval fail differently, how to fuse them with Reciprocal Rank Fusion instead of hand-tuned weights, what a cross-encoder reranker actually does and what it costs in latency, and how to size a retrieve-then-rerank pipeline and measure each stage separately. Tue, 11 Aug 2026 00:00:00 GMT https://pavlo.sh/blog/why-ai-agent-pilots-never-reach-production https://pavlo.sh/blog/why-ai-agent-pilots-never-reach-production Three independent studies in 2026 put the enterprise AI agent pilot failure rate somewhere between 86 and 89 percent. The interesting part is that almost none of those failures are model failures. This article breaks down where pilots actually die, why a great demo is a poor predictor of a working system, how per-step reliability compounds into production failure, and what the small minority of teams that ship are doing differently. Sun, 02 Aug 2026 00:00:00 GMT https://pavlo.sh/blog/eu-ai-act-august-2026-what-developers-actually-have-to-ship https://pavlo.sh/blog/eu-ai-act-august-2026-what-developers-actually-have-to-ship Most engineering teams read the headline that the EU AI Act was delayed and stopped paying attention. The delay applies to high-risk systems, not to Article 50, which becomes enforceable on 2 August 2026 with fines up to 15 million euro or 3 percent of global turnover. If you ship a chatbot, an agent, or anything that generates text, images, audio, or video, this is the deadline that lands on your backlog. This guide translates Article 50 into engineering requirements: what to disclose, what to mark machine-readable, where the exceptions actually apply, and what to build this week. Sat, 25 Jul 2026 00:00:00 GMT https://pavlo.sh/blog/claude-vs-gpt-vs-gemini-how-to-choose-in-2026 https://pavlo.sh/blog/claude-vs-gpt-vs-gemini-how-to-choose-in-2026 Anthropic, OpenAI, and Google are all shipping new flagship models every few weeks in 2026, and none of them wins on every axis anymore. This guide compares the current Claude, GPT-5.6, and Gemini 3 lineups model by model, with real pricing, real benchmark data, and a practical framework for choosing the right model, or the right combination of models, for coding, research, enterprise work, and cost-sensitive pipelines. Wed, 15 Jul 2026 00:00:00 GMT https://pavlo.sh/blog/prompt-caching-and-semantic-caching-cut-llm-costs https://pavlo.sh/blog/prompt-caching-and-semantic-caching-cut-llm-costs Two techniques share the word "cache" in LLM engineering, and teams mix them up constantly, even though they solve different problems at different layers. This guide explains how prompt caching reuses already-processed tokens to cut input costs, how semantic caching skips the model entirely on near-duplicate queries, the real numbers from Anthropic, OpenAI, and Gemini, and the failure modes, like stale answers and false positives, that make semantic caching far riskier than it looks. Sat, 04 Jul 2026 00:00:00 GMT https://pavlo.sh/blog/chunking-strategies-for-rag-the-decision-that-makes-or-breaks-retrieval https://pavlo.sh/blog/chunking-strategies-for-rag-the-decision-that-makes-or-breaks-retrieval Chunking is the quietest, most consequential decision in any RAG system. The chunk is the unit of retrieval, so if you split your documents badly, no embedding model or reranker can save you. This guide explains the real trade-off between precision and context, the chunk sizes and overlaps that actually work, and the full ladder of strategies, from recursive splitting to late chunking and contextual retrieval, so you know which one your documents actually need. Tue, 30 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/model-context-protocol-mcp-explained https://pavlo.sh/blog/model-context-protocol-mcp-explained Every AI agent that connects to a tool, a database, or an API used to need its own custom integration code. Model Context Protocol replaces that with one standard. This guide explains what MCP actually is, how hosts, clients, and servers fit together, how tools, resources, and prompts differ, why it spread across the industry so fast, and the real security and token-cost tradeoffs you take on the moment you connect a server. Sat, 27 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/rag-vs-fine-tuning-vs-prompting-how-to-choose https://pavlo.sh/blog/rag-vs-fine-tuning-vs-prompting-how-to-choose The most common architecture question in LLM applications has no one-size answer. This guide explains what prompting, RAG, and fine-tuning each actually change, why the difference is knowledge versus behavior, why most teams reach for fine-tuning too early, and how to walk the 2026 decision ladder—prompt, then RAG, then fine-tune, then distill—without wasting months and budget. Wed, 24 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/how-to-integrate-artificial-intelligence-into-business-processes-without-disrupting-everything https://pavlo.sh/blog/how-to-integrate-artificial-intelligence-into-business-processes-without-disrupting-everything A practical, updated playbook for integrating AI into business processes in the LLM era. Learn how to assess readiness, pick high-impact use cases, decide between traditional ML and LLMs, evaluate systems before they ship, manage security risks like prompt injection, and roll out gradually without breaking operations. Sat, 13 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/optimization-of-ml-models-advanced-techniques-to-reduce-resource-consumption https://pavlo.sh/blog/optimization-of-ml-models-advanced-techniques-to-reduce-resource-consumption Models keep getting bigger while budgets do not. This updated guide covers the techniques that make machine learning and LLMs cheaper to run without sacrificing accuracy: quantization, pruning, knowledge distillation, feature selection, efficient hardware, and adaptive computation, with practical notes on applying them to large language models in production. Sat, 13 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/embeddings-explained-choosing-the-right-model-and-vector-database-for-production https://pavlo.sh/blog/embeddings-explained-choosing-the-right-model-and-vector-database-for-production Your RAG system is only as good as the embeddings underneath it. This guide explains what embeddings actually are, how to choose an embedding model in 2026 without trusting leaderboards blindly, how dimensions affect cost and latency, and how to pick a vector database (Pinecone, Qdrant, Weaviate, Milvus, pgvector) based on scale, indexing, and filtering rather than hype. Sun, 07 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/prompt-injection-security-holes-in-every-llm-app https://pavlo.sh/blog/prompt-injection-security-holes-in-every-llm-app Prompt injection is the number one security risk in LLM applications, and there is no patch that makes it go away. This guide explains direct and indirect injection, how data gets exfiltrated through tools and markdown images, the lethal trifecta that makes agents dangerous, and the defense-in-depth strategy that actually reduces your blast radius in production. Tue, 02 Jun 2026 00:00:00 GMT https://pavlo.sh/blog/llm-evaluation-why-your-demo-works-but-production-fails https://pavlo.sh/blog/llm-evaluation-why-your-demo-works-but-production-fails Most LLM applications demo perfectly and then break with real users. This guide explains how to evaluate LLM applications properly: how to build an eval dataset, the metrics that actually matter, how to use LLM-as-a-judge without fooling yourself, and how to catch regressions before your users do. Sat, 30 May 2026 00:00:00 GMT https://pavlo.sh/blog/why-tokens-matter https://pavlo.sh/blog/why-tokens-matter Tokens are the fundamental unit of everything you do with LLMs: pricing, context limits, latency, retrieval, even multilingual fairness. This article explains what tokens really are, why they behave strangely across languages, and how a practical understanding of tokenization changes the way you design AI systems. Sat, 25 Apr 2026 00:00:00 GMT https://pavlo.sh/blog/rag-in-production-what-nobody-tells-you https://pavlo.sh/blog/rag-in-production-what-nobody-tells-you RAG sounds simple in theory: retrieve relevant chunks, inject them into the prompt, get better answers. In production, the reality is far messier. This guide covers the real failure modes, including chunking pitfalls, embedding drift, retrieval quality collapse, and latency traps, and what actually works to fix them. Tue, 21 Apr 2026 00:00:00 GMT https://pavlo.sh/blog/llm-context-window-limitations-accuracy-degradation https://pavlo.sh/blog/llm-context-window-limitations-accuracy-degradation Large language models advertise million-token context windows, but longer inputs silently degrade accuracy. Learn why the "lost in the middle" problem affects every major LLM, and what RAG and prompt structuring strategies actually work for production AI systems. Thu, 16 Apr 2026 00:00:00 GMT https://pavlo.sh/blog/agent-skills-vs-multi-agent-systems-architectural-shift-ai https://pavlo.sh/blog/agent-skills-vs-multi-agent-systems-architectural-shift-ai Agent Skills introduce a new paradigm for building AI systems by packaging operational knowledge into reusable modules. But can they truly replace multi-agent architectures? This article explores the trade-offs, strengths, and future of both approaches. Tue, 17 Mar 2026 00:00:00 GMT https://pavlo.sh/blog/openai-agents-sdk-build-agentic-ai-apps-in-python-with-ease https://pavlo.sh/blog/openai-agents-sdk-build-agentic-ai-apps-in-python-with-ease Learn how to build powerful, customizable agentic AI applications in Python using the OpenAI Agents SDK. Discover multi-agent orchestration, guardrails, and built-in tracing for production-ready AI workflows. Sat, 31 May 2025 00:00:00 GMT