AI Engineering: SDLC, contesto e ingegneria dell'inferenza di LLM locali

Traduzione in italiano…

🚀 **BUILD IT CLUB presents:** **From enterprise AI adoption to efficient agents and self-hosted inference.** Join us for another exciting **AI & Tech Meetup in Kraków**, where three speakers will share practical experiences building, integrating, and running AI systems in the real world. 📅 **Monday, November 16, 2026** 🕡 **6:45 PM** 📍 **Grand Parade, Kotlarska 11, 31-539 Kraków** 🎟️ **Free event \| Everyone is welcome\!** AI is changing how we build software, but moving beyond prototypes introduces challenges that go far beyond prompting a language model. How do organizations integrate AI into their software development lifecycle? How can AI agents become more efficient without wasting thousands of tokens? And what does it actually take to deploy a powerful open-source LLM and build your own coding agent? This November, we're bringing together three perspectives on the engineering behind modern AI systems. ## 🎤 Speaker 1: Iryna Marchenko **Data Architect @ Grand Parade** LinkedIn: [https://www.linkedin.com/in/irynamarchenko/](https://www.linkedin.com/in/irynamarchenko/) ### AI Meets Our SDLC: What the Pilot Is Teaching Us Introducing AI across a complex organization exposes gaps in processes, ownership, and documentation. In this session, Iryna will share how an ongoing AI SDLC pilot is helping address these challenges through shared guardrails, established standards, reusable skills, and human review. We'll explore how process challenges connect with the technical framework, and how evidence, context, requirements, and design artifacts can form a traceable chain throughout software development. The current pilot focuses on documentation and design, with implementation and QA planned for the next phase. **What you'll learn:** * The organizational and technical challenges of adopting AI in software development. * How standards, reusable skills, and human oversight can support AI-assisted workflows. * Why traceability between requirements, context, and design matters. * Early lessons from an AI SDLC pilot and the challenges that remain unresolved. ## 🎤 Speaker 2: Weronika Kołodziej **Developer Advocate @ RavenDB** LinkedIn: [https://www.linkedin.com/in/weronikakol/](https://www.linkedin.com/in/weronikakol/) ### Context Is a Resource: How to Burn Fewer Tokens and Hit Zero Retrieval Egress LLM agents are introducing a new kind of workload into our infrastructure. They may not perform much computation themselves, but they move significant amounts of data. Every request sends context to a model, creating costs in tokens, network traffic, and infrastructure. Using an open-source hidden-city flight agent running on Kubernetes, Weronika will trace a single question through the system and show how an inefficient architecture can send **40,000 tokens per request**, and how that was reduced to approximately **2,500 tokens**. The session explores the evolution of agent architectures, from placing everything directly into prompts to complex RAG pipelines, before simplifying the design through tool calling, push-based communication instead of polling, and a database running inside the cluster. **What you'll learn:** * Why context management matters for AI agent performance and cost. * How to reduce token usage without unnecessarily sacrificing functionality. * How architectural decisions affect retrieval traffic and network egress. * How to measure tokens and bytes per conversation turn. * Why AI agents need context budgets, just as Kubernetes workloads need resource limits. ## 🎤 Speaker 3: Safari Germain **Lead Machine Learning Engineer @ Admind Branding & Communication** LinkedIn: [https://www.linkedin.com/in/safari-germain/](https://www.linkedin.com/in/safari-germain/) ### Inference Engineering in Practice: From Local LLM Deployment to Your Own Coding Agent Everyone is building applications with LLM APIs. But what happens when you want to run the model yourself? In this hands-on engineering session, Germain will walk through his journey of deploying an open-source language model on local hardware and using it to power a custom AI coding assistant. Starting with **Qwen3.6-35B-A3B NVFP4**, a quantized mixture-of-experts model, we'll explore how to serve a large language model using **vLLM inside Docker**, download and manage model weights from Hugging Face, and expose the model through an API that a coding agent can interact with. From there, we'll connect the inference server to a Python-based agent engine and renderer, demonstrating how a locally hosted model can power a practical coding assistant without relying on a commercial hosted LLM API for inference. But running an LLM is only half the story. We'll also look at the engineering challenges behind inference: model loading, GPU memory constraints, quantization, context length, response latency, and measuring where time is actually spent between the model and the application. The goal is to show what happens behind the scenes when you move from simply consuming an LLM API to **owning and operating the model that powers your application**. **What you'll learn:** * How to deploy an open-source LLM locally using Docker and vLLM. * How quantization helps reduce model memory requirements. * How to connect a self-hosted LLM to a custom Python coding agent. * How the model server, agent engine, and user-facing interface work together. * What affects inference performance, including latency and GPU memory. * Practical lessons from building and operating your own local AI development tool. ## 👥 Who should attend? This meetup is for **software engineers, ML/AI engineers, data engineers, cloud and platform architects, DevOps/MLOps practitioners, technical leaders**, and anyone curious about how AI systems work beyond the demo. You don't need to be an AI expert. Bring your curiosity, technical questions, and willingness to learn! ## 🗓️ Event schedule * **6:45 PM:** Arrival, registration & networking * **7:00 PM:** Iryna Marchenko \| AI in the SDLC **Break at 7:30PM** * **7:35 PM:** Weronika Kołodziej \| Context Engineering **Break at 8:05PM** * **8:10 PM:** Safari Germain \| Local LLMs & Coding Agents * **8:40 PM onwards:** Open discussions & networking Each presentation will include time for questions. ## 🤝 Hosted at Grand Parade Kraków A big thank you to **Grand Parade** for welcoming our community and supporting knowledge sharing among engineers in Kraków. Build It Club is a community where engineers, researchers, and technology enthusiasts come together to exchange practical knowledge, discover opportunities, and build meaningful connections. 💡 **Have something interesting to share?** We're always looking for speakers, collaborators, and organizations interested in contributing to our community. Join our community: [https://www.meetup.com/build-it-club](https://www.meetup.com/build-it-club) **Come for the technology. Stay for the conversations and connections. 🚀**