# Helain Zimmermann - AI & ML Engineering Blog

> Technical blog about AI, automation, RAG systems, NLP, privacy-preserving AI, and applied mathematics by Helain Zimmermann, Co-Founder & CTO at Ailog.

## About the Author
Helain Zimmermann is an AI Engineer and Co-Founder & CTO at Ailog, specializing in RAG systems, privacy-preserving NLP, and machine learning. Alumni of ENSIMAG and KTH Royal Institute of Technology.

## Blog Articles

### AI & ML

- [GPT-5.4 vs Gemini 3.1 Ultra vs Claude Opus 4.6: Picking the Right Frontier Model](https://helain-zimmermann.com/blog/gpt-5-4-vs-gemini-3-1-ultra-vs-claude-opus-4-6-picking-the-right-frontier-model): A benchmark-driven comparison of GPT-5.4, Gemini 3.1 Ultra, and Claude Opus 4.6 with practical recommendations by use case
- [TurboQuant: How Google's 3-Bit KV Cache Compression Cuts LLM Memory by 6x](https://helain-zimmermann.com/blog/turboquant-google-3-bit-kv-cache-compression-for-llm-inference): Google's TurboQuant compresses KV cache to 3 bits with zero accuracy loss using rotation-based quantization, enabling 6x memory savings for LLM inference
- [Open-Source LLMs in 2026: DeepSeek V3.2 vs Llama vs Mistral](https://helain-zimmermann.com/blog/open-source-llms-2026-deepseek-vs-llama-vs-mistral): A practical comparison of the leading open-source language models with benchmarks, deployment costs, and use case recommendations
- [Anthropic Mythos and the Next-Gen Reasoning Race](https://helain-zimmermann.com/blog/anthropic-mythos-and-the-next-gen-reasoning-race): What the leaked Anthropic Mythos model reveals about the frontier reasoning arms race and what developers should prepare for
- [Apple's Siri Rebuild: What Gemini Integration Tells Us About On-Device AI](https://helain-zimmermann.com/blog/apple-siri-gemini-what-the-biggest-consumer-ai-rebuild-means): Apple is rebuilding Siri with Google's Gemini models for reasoning and on-screen awareness, shipping with iOS 26.4 in 2026
- [Gemini 3.1 Flash-Lite and the Arrival of Sub-Dollar Inference](https://helain-zimmermann.com/blog/gemini-3-1-flash-lite-and-the-arrival-of-sub-dollar-inference): Google's Gemini 3.1 Flash-Lite costs $0.25 per million input tokens while outperforming models at 4x the price point
- [GliNER2: Unified Entity and Relation Extraction in One Framework](https://helain-zimmermann.com/blog/gliner2-unified-entity-and-relation-extraction-in-one-framework): GliNER2 merges NER, relation extraction, text classification, and structured data extraction into a single schema-driven inference call
- [GPT-5.4 Tool Search: How Dynamic Discovery Reshapes Agent Design](https://helain-zimmermann.com/blog/gpt-5-4-tool-search-how-dynamic-discovery-reshapes-agent-design): GPT-5.4 introduces Tool Search for dynamic tool discovery, cutting token usage by 47% while maintaining accuracy across large tool inventories
- [LTX 2.3: What Open-Source 4K Video Generation Means for AI Engineers](https://helain-zimmermann.com/blog/ltx-2-3-open-source-4k-video-generation-for-ai-engineers): Lightricks ships LTX 2.3, a 22B parameter open-source model generating 4K video with synchronized audio under Apache 2.0
- [The Million-Token Context Window: What Actually Changes](https://helain-zimmermann.com/blog/million-token-context-windows-what-actually-changes): GPT-5.4, Gemini 3.1, and Nemotron 3 all ship with 1M token contexts, but using them effectively requires rethinking retrieval and architecture
- [Mistral Small 4: One MoE to Replace Three Models](https://helain-zimmermann.com/blog/mistral-small-4-one-moe-to-replace-three-models): Mistral Small 4 unifies instruct, reasoning, and multimodal workloads into a single 119B parameter MoE model under Apache 2.0
- [Nemotron 3 Super: The Hybrid Mamba-Transformer Built for Agentic Coding](https://helain-zimmermann.com/blog/nemotron-3-super-hybrid-mamba-transformer-for-agentic-coding): NVIDIA's Nemotron 3 Super combines Mamba and Transformer layers in a 120B MoE model scoring 60.47% on SWE-Bench Verified
- [Sub-10B Models Are Winning the Efficiency Race](https://helain-zimmermann.com/blog/sub-10b-models-winning-the-efficiency-race): How models under 10 billion parameters are outperforming much larger rivals through distillation, sparse architectures, and smarter training
- [World Models: The NLP Paradigm Beyond Next-Token Prediction](https://helain-zimmermann.com/blog/world-models-the-nlp-paradigm-beyond-next-token-prediction): World models build internal representations of environments to enable causal reasoning, continuity, and grounded understanding in NLP systems
- [Adult Neurogenesis: How Your Brain Keeps Growing New Neurons](https://helain-zimmermann.com/blog/adult-neurogenesis-how-your-brain-keeps-growing-new-neurons): Discover how your adult brain grows new neurons, what recent research reveals about neuroplasticity, and why it matters for AI engineers
- [Neuromorphic Computing Meets Neurogenesis: Inspiring Plasticity in AI](https://helain-zimmermann.com/blog/neuromorphic-computing-meets-neurogenesis-inspiring-plasticity-in-ai): How biological neurogenesis and neuromorphic hardware are inspiring new approaches to plasticity, lifelong learning, and catastrophic forgetting in AI
- [GLM-5, Kimi K2.5, DeepSeek V3: A 2026 Panorama of Frontier Chinese LLMs](https://helain-zimmermann.com/blog/glm5-kimi-deepseek-panorama-2026-chinese-llms): A comparative overview of the leading Chinese language models: architecture, hardware, benchmarks, licensing, and how to test them from Europe.
- [Why Chinese AI Models Are Breaking the Economics of AI](https://helain-zimmermann.com/blog/why-chinese-models-are-breaking-ai-economics): With benchmark scores rivaling GPT-4 at a fraction of the cost, Chinese LLMs are reshaping how startups and developers choose their AI models.
- [Chinese AI Models Are Catching Up, And Sometimes Surpassing, The West](https://helain-zimmermann.com/blog/chinese-ai-models-catching-up-and-surpassing-the-west): A beginner's guide to Kimi K2, GLM-5, Qwen, and DeepSeek: the Chinese large language models that rival GPT-4 and why they matter.
- [Open Source AI vs Closed AI: Why It Matters More Than Ever](https://helain-zimmermann.com/blog/open-source-vs-closed-ai-why-everyone-is-talking-about-it): Understanding the difference between open-weights models and closed APIs, and why this debate is reshaping the AI industry in 2026.
- [Building Custom Tokenizers for Domain-Specific NLP](https://helain-zimmermann.com/blog/building-custom-tokenizers-for-domain-specific-nlp): Learn how to design, implement, and evaluate custom tokenizers for domain-specific NLP, with practical Python examples and RAG-focused guidance.
- [Embedding Models Compared: OpenAI vs Open-Source](https://helain-zimmermann.com/blog/embedding-models-compared-openai-vs-open-source): Compare OpenAI and open-source embedding models for RAG, search, and clustering, with tradeoffs, benchmarks, costs, and practical code examples.
- [Multimodal AI: Combining Vision and Language Models](https://helain-zimmermann.com/blog/multimodal-ai-combining-vision-and-language-models): Learn how to build practical multimodal AI systems that combine vision and language models, from architectures to PyTorch and CLIP code examples.
- [Named Entity Recognition with Modern NLP](https://helain-zimmermann.com/blog/named-entity-recognition-with-modern-nlp): Learn modern Named Entity Recognition, from classical CRFs to transformer-based models, practical pipelines, privacy, evaluation, and production tips.
- [Understanding Transformer Architectures](https://helain-zimmermann.com/blog/understanding-transformer-architectures): Deep dive into transformer architectures, from self-attention math to practical variants for RAG, privacy NLP, and production systems.
- [Understanding Vector Databases: A Beginner's Guide](https://helain-zimmermann.com/blog/understanding-vector-databases): Learn what vector databases are, how they work under the hood, and which one to choose for your AI application.

### RAG Systems

- [GraphRAG in Practice: When Vector Search Alone Is Not Enough](https://helain-zimmermann.com/blog/graphrag-in-practice-when-vector-search-alone-isnt-enough): How combining knowledge graphs with vector retrieval achieves up to 99% search precision and when you should adopt GraphRAG
- [RAG for Code: Building Retrieval Systems Over Codebases](https://helain-zimmermann.com/blog/rag-for-code-retrieval-over-codebases): How to adapt RAG techniques for code search with syntax-aware chunking, code embeddings, and AST-based retrieval
- [Enterprise RAG with Citation Tracking and Audit Trails](https://helain-zimmermann.com/blog/enterprise-rag-with-citation-tracking-and-audit-trails): How to build RAG systems that trace every answer back to its source document for compliance, trust, and debugging
- [2026: The Year of AI Memory Beyond Basic RAG](https://helain-zimmermann.com/blog/2026-the-year-of-ai-memory-beyond-basic-rag): How AI memory systems are evolving past basic RAG with episodic, semantic, and procedural memory for persistent, context-aware agents
- [Multimodal RAG 2026: Vision and Text for State-of-the-Art Pipelines](https://helain-zimmermann.com/blog/multimodal-rag-2026-vision-and-text-for-state-of-the-art-pipelines): Build production multimodal RAG pipelines combining vision and text retrieval with Qwen3-VL, cross-modal fusion, and cost optimization strategies.
- [Agentic RAG: The Next Evolution](https://helain-zimmermann.com/blog/agentic-rag-the-next-evolution): Explore Agentic RAG, where LLM agents plan, search, and verify across tools. Design patterns, code, and pitfalls for production-ready systems.
- [Chunking Strategies for RAG Pipelines](https://helain-zimmermann.com/blog/chunking-strategies-for-rag-pipelines): Learn practical chunking strategies for RAG pipelines, from basic splits to adaptive and hybrid methods, with code and evaluation tips.
- [Evaluating RAG System Performance](https://helain-zimmermann.com/blog/evaluating-rag-system-performance): Learn how to rigorously evaluate Retrieval-Augmented Generation systems with practical metrics, tooling, and Python examples for production setups.
- [Hybrid Search: Combining Dense and Sparse Retrieval](https://helain-zimmermann.com/blog/hybrid-search-combining-dense-and-sparse-retrieval): Learn how to design and implement hybrid search that combines dense and sparse retrieval, with practical patterns, tradeoffs, and Python code examples.
- [Knowledge Graphs Meet LLMs: Structured RAG Architectures](https://helain-zimmermann.com/blog/knowledge-graphs-meet-llms-structured-rag-architectures): How to combine knowledge graphs with LLMs for structured RAG architectures, with patterns, code, and tradeoffs for production systems.
- [Retrieval-Augmented Generation: A Complete Guide](https://helain-zimmermann.com/blog/retrieval-augmented-generation-a-complete-guide): Beginner-friendly guide to Retrieval-Augmented Generation, with architecture, tradeoffs, vector DBs, privacy tips, and Python code examples.
- [Scaling RAG Systems to Millions of Documents](https://helain-zimmermann.com/blog/scaling-rag-systems-to-millions-of-documents): Learn practical strategies to scale Retrieval-Augmented Generation systems to millions of documents, from indexing and storage to latency and cost tuning.
- [Semantic Search vs Keyword Search: When to Use What](https://helain-zimmermann.com/blog/semantic-search-vs-keyword-search-when-to-use-what): Understand the differences between keyword and semantic search, when to use each, and how to implement basic semantic search in Python.
- [Vector Database Performance Benchmarks](https://helain-zimmermann.com/blog/vector-database-performance-benchmarks): Learn how to design, run, and interpret vector database performance benchmarks for real-world RAG systems, with code, metrics, and pitfalls.
- [Building Production-Ready RAG Systems](https://helain-zimmermann.com/blog/building-production-ready-rag-systems): A practical guide to designing and deploying Retrieval-Augmented Generation systems that scale, from chunking strategies to vector store optimization.

### AI Agents

- [Multi-Agent Systems in Production: Patterns That Actually Work](https://helain-zimmermann.com/blog/multi-agent-systems-in-production-patterns-that-work): Proven architectural patterns for deploying multi-agent AI systems that survive contact with real workloads and real users
- [The Agentic AI Foundation: Open Governance for Agent Protocols](https://helain-zimmermann.com/blog/the-agentic-ai-foundation-open-governance-for-agent-protocols): How the Linux Foundation initiative with OpenAI, Anthropic, Google, Microsoft, AWS, and Block is shaping the future of AI agent standards
- [Domain-Specific AI Agents: Lessons from Siemens Fuse](https://helain-zimmermann.com/blog/domain-specific-ai-agents-lessons-from-siemens-fuse): What Siemens' EDA AI Agent system teaches us about building purpose-built autonomous agents for specialized industries
- [MCP vs A2A: Picking the Right Agent Communication Protocol](https://helain-zimmermann.com/blog/mcp-vs-a2a-picking-the-right-agent-protocol): A practical comparison of Model Context Protocol and Agent-to-Agent protocol to help you choose the right standard for your AI system
- [AI and Finance: How Autonomous Agents Are Transforming Trading](https://helain-zimmermann.com/blog/ai-and-finance-how-autonomous-agents-are-transforming-trading): How autonomous AI agents with their own wallets are reshaping trading and DeFi, from market making to risk management and ethical concerns
- [Multi-Agent Multi-LLM Architectures: A 2026 Guide](https://helain-zimmermann.com/blog/multi-agent-multi-llm-architectures-a-2026-guide): A practical guide to multi-agent multi-LLM architectures covering agent roles, communication patterns, MoE, RLVR, and distributed scalability
- [The Multi-Agent Systems Explosion: 327% Adoption Growth in 2026](https://helain-zimmermann.com/blog/the-multi-agent-systems-explosion-327-adoption-growth-in-2026): Why multi-agent AI systems are seeing 327% adoption growth in 2026 and what it means for startups, enterprises, and the future of automation
- [AI Trends 2026: The Agentic Revolution and Hybrid Architectures](https://helain-zimmermann.com/blog/ai-trends-2026-the-agentic-revolution-and-hybrid-architectures): Comprehensive synthesis of 2026 AI trends from agent swarms and sovereign infrastructure to hybrid architectures and RLVR breakthroughs
- [FinAgent and Beyond: Multimodal Foundation Agents for Trading](https://helain-zimmermann.com/blog/finagent-and-beyond-multimodal-foundation-agents-for-trading): Deep dive into FinAgent's multimodal architecture for financial trading, covering dual-level reflection, memory retrieval, and agent swarm extensions
- [Multimodal AI Agents: When AI Sees, Hears, and Acts](https://helain-zimmermann.com/blog/multimodal-ai-agents-when-ai-sees-hears-and-acts): Discover how multimodal AI agents combine text, images, and audio to see, hear, and act. Practical examples from visual RAG to accessibility.
- [Agents + Open Weights: Toward a Linux of AI Agents, Or a Security Nightmare?](https://helain-zimmermann.com/blog/agents-plus-open-weights-linux-of-agents-or-security-nightmare): Combining open-source AI agents like OpenClaw with frontier open-weight models creates unprecedented opportunities and risks for the AI ecosystem.
- [OpenClaw Explained: What Is the AI Agent Everyone Is Talking About?](https://helain-zimmermann.com/blog/openclaw-explained-for-beginners): A beginner-friendly guide to OpenClaw, the open-source AI agent that can browse the web, send messages, and automate tasks, and why it matters.
- [Model Context Protocol: Connecting LLMs to External Tools](https://helain-zimmermann.com/blog/model-context-protocol-connecting-llms-to-external-tools): Learn how Model Context Protocol connects LLMs to tools, RAG systems, and services, with Python examples and practical integration patterns.
- [Building a Multi-Agent AI System](https://helain-zimmermann.com/blog/building-a-multi-agent-ai-system): Learn how to design, coordinate, and deploy robust multi-agent AI systems, from architecture and tools to failure modes and production concerns.

### Engineering

- [LLMOps: Managing LLM-Powered Production Systems](https://helain-zimmermann.com/blog/llmops-managing-llm-powered-production-systems): How LLMOps extends traditional MLOps with prompt versioning, guardrails, retrieval monitoring, and cost optimization
- [On-Device NLP: Running Language Models at the Edge with TinyML](https://helain-zimmermann.com/blog/on-device-nlp-running-language-models-at-the-edge-with-tinyml): How model compression techniques like quantization, pruning, and distillation enable NLP inference on edge devices without cloud dependencies
- [End-to-End Multi-Agent Systems: Design Patterns from IEEE CAI 2026](https://helain-zimmermann.com/blog/end-to-end-multi-agent-systems-design-patterns-from-ieee-cai-2026): Design patterns for production multi-agent systems from IEEE CAI 2026 covering planning, execution, fault tolerance, and scaling
- [Chinese AI and Sovereign Hardware: GLM-5 on Huawei Ascend](https://helain-zimmermann.com/blog/chinese-ai-sovereign-hardware-glm5-huawei-ascend): How Zhipu trained a frontier language model entirely on domestic Chinese chips, and what this means for AI geopolitics, sanctions, and infrastructure independence.
- [OpenClaw as a Case Study in Autonomous Agent Attack Surfaces](https://helain-zimmermann.com/blog/openclaw-attack-surface-case-study-autonomous-agents): A technical threat model analysis of AI agents that can act in the real world: network exposure, API key management, extension supply chains, and memory compromise.
- [Anatomy of an OpenClaw Agent: Architecture, Integrations, and Limits](https://helain-zimmermann.com/blog/anatomy-of-an-openclaw-agent): A semi-technical deep dive into how OpenClaw connects to messaging platforms, LLMs, and third-party APIs, and where it falls short.
- [Building Real-Time ML Inference Pipelines](https://helain-zimmermann.com/blog/building-real-time-ml-inference-pipelines): Learn how to design and implement real-time ML inference pipelines, from architecture and latency budgets to queues, batching, monitoring and Python code.
- [CI/CD Pipelines for Machine Learning Projects](https://helain-zimmermann.com/blog/cicd-pipelines-for-machine-learning-projects): Learn how to design practical CI/CD pipelines for ML projects, covering testing, data checks, model evaluation, deployment and MLOps tooling.
- [Deploying ML Models with FastAPI and Docker](https://helain-zimmermann.com/blog/deploying-ml-models-with-fastapi-and-docker): Learn how to containerize and deploy ML models using FastAPI and Docker, with patterns for scaling, performance, and production-ready setups.
- [Monitoring ML Models in Production](https://helain-zimmermann.com/blog/monitoring-ml-models-in-production): Practical guide to monitoring ML models in production, covering metrics, drift, data quality, logging, alerts, and code patterns in Python.
- [Python Best Practices for ML Engineers](https://helain-zimmermann.com/blog/python-best-practices-for-ml-engineers): Practical Python best practices for ML engineers, from project structure and typing to performance, testing, and production-ready code.

### AI Security

- [Zero-Trust for AI Agents: Securing Non-Human Identities](https://helain-zimmermann.com/blog/zero-trust-for-ai-agents-securing-non-human-identities): How to apply zero-trust security principles to autonomous AI agents that outnumber human users 100 to 1
- [EU AI Act 2026: What Developers Need to Comply with Now](https://helain-zimmermann.com/blog/eu-ai-act-2026-what-developers-need-to-comply-with-now): A practical compliance checklist for AI developers as EU AI Act enforcement moves from draft to reality in 2026
- [AI Agents in Finance: MNPI Risks and Cross-Deal Contamination](https://helain-zimmermann.com/blog/ai-agents-in-finance-mnpi-risks-and-cross-deal-contamination): How RAG systems in finance can leak MNPI across deals, and practical mitigation strategies for quants building compliant AI pipelines
- [AI Alignment and Safety in a Multipolar World](https://helain-zimmermann.com/blog/ai-alignment-safety-in-a-multipolar-world): As the US, China, and the open-source community each push frontier AI forward, can we agree on safety standards, or are we heading for fragmented norms?
- [AI Regulation: How Open Source Is Being Instrumentalized](https://helain-zimmermann.com/blog/ai-regulation-how-open-source-is-being-instrumentalized): From the EU AI Act to global policy debates, open source is positioned as both a tool for accountability and a shield for concentrated power.
- [AI and Security: How Agents Like OpenClaw Can Be Exploited](https://helain-zimmermann.com/blog/ai-security-how-agents-like-openclaw-can-be-exploited): Exposed instances, stolen API keys, and malicious extensions: how autonomous AI agents create new attack vectors and what you can do about it.
- [Data Privacy in the Age of Large Language Models](https://helain-zimmermann.com/blog/data-privacy-in-the-age-of-large-language-models): Practical strategies to protect data privacy in LLM workflows, from architecture and redaction to logs, RAG, and compliant deployment patterns.
- [Federated Learning for Privacy-Preserving AI](https://helain-zimmermann.com/blog/federated-learning-for-privacy-preserving-ai): Learn how to design and ship federated learning systems for privacy-preserving AI, from protocols and architectures to practical Python examples.
- [Introduction to Differential Privacy for NLP](https://helain-zimmermann.com/blog/introduction-to-differential-privacy-for-nlp): Advanced introduction to differential privacy for NLP practitioners, with practical Python examples, tradeoffs, and system design advice.
- [Privacy-Preserving NLP: Protecting Sensitive Data in Language Models](https://helain-zimmermann.com/blog/privacy-preserving-nlp): Exploring techniques for mitigating memorization risks in LLMs, from differential privacy to anonymization, based on research at INRIA.

### Getting Started

- [Running Local LLMs with Ollama: A Practical Guide](https://helain-zimmermann.com/blog/running-local-llms-with-ollama-a-practical-guide): How to set up, run, and integrate local language models using Ollama for development, testing, and privacy-sensitive workloads
- [Prompt Engineering for Multi-Agent Workflows](https://helain-zimmermann.com/blog/prompt-engineering-for-multi-agent-workflows): Techniques for writing effective prompts that coordinate multiple AI agents working together on complex tasks
- [Building Your First MCP Server in Python](https://helain-zimmermann.com/blog/building-your-first-mcp-server-in-python): A step-by-step tutorial to create a Model Context Protocol server that exposes custom tools to AI agents
- [Building a RAG Chatbot from Scratch with Python](https://helain-zimmermann.com/blog/building-a-rag-chatbot-from-scratch-with-python): Learn how to build a Retrieval-Augmented Generation (RAG) chatbot from scratch in Python, from data loading to retrieval and LLM integration.
- [Fine-Tuning Open-Source LLMs with LoRA and QLoRA](https://helain-zimmermann.com/blog/fine-tuning-open-source-llms-with-lora-and-qlora): Learn how to fine-tune open-source LLMs efficiently using LoRA and QLoRA, with practical code, tips, and trade-offs for production systems.
- [Building AI Agents with LangChain and LangGraph](https://helain-zimmermann.com/blog/building-ai-agents-with-langchain-and-langgraph): Learn how to build robust AI agents with LangChain and LangGraph, from simple tool calls to multi-step workflows, with practical Python examples.
- [Getting Started with PyTorch for Deep Learning](https://helain-zimmermann.com/blog/getting-started-with-pytorch-for-deep-learning): Beginner friendly guide to getting started with PyTorch, from tensors to training your first neural network, with practical Python examples.
- [LLM Evaluation Frameworks: Beyond Perplexity](https://helain-zimmermann.com/blog/llm-evaluation-frameworks-beyond-perplexity): Go beyond perplexity with practical LLM evaluation: task metrics, judge models, rubrics, RAG-specific checks, and production feedback loops.
- [Prompt Engineering Best Practices for Production](https://helain-zimmermann.com/blog/prompt-engineering-best-practices-for-production): Practical prompt engineering best practices for production systems including structure, evaluation, safety, RAG integration and maintainability.
- [Fine-Tuning LLMs on Custom Data: A Practical Guide](https://helain-zimmermann.com/blog/fine-tuning-llms-custom-data): When to fine-tune vs use RAG, how to prepare your data, and a step-by-step guide to LoRA fine-tuning with Hugging Face Transformers.

### Industry

- [AI in Healthcare 2026: The Documentation Time Revolution](https://helain-zimmermann.com/blog/ai-in-healthcare-2026-the-documentation-time-revolution): How AI clinical assistants are cutting documentation time by 42% and what healthcare organizations should learn from early adopters

### Strategy

- [The AI Adoption ROI Gap: Why 88% See Returns but Only 15% See Profit](https://helain-zimmermann.com/blog/the-ai-adoption-roi-gap-why-88-percent-see-returns-but-only-15-percent-profit): Analyzing the disconnect between AI investment optimism and actual profitability, with lessons from successful enterprise deployments

## Contact
- Website: https://helain-zimmermann.com
- LinkedIn: https://www.linkedin.com/in/helain-zimmermann/
- GitHub: https://github.com/L1ZGitHub