Paul Iusztin

Paul Iusztin is the founder and chief executive of Decoding AI, an AI engineer and educator, and coauthor of the LLM Engineer’s Handbook. He builds and teaches production-ready AI systems that turn personal knowledge into reusable agent memory, preserve human judgment in generated writing, and evaluate outputs against real-world requirements.

Raised in Timișoara, Romania, Iusztin began tutoring mathematics and programming in 2016 and working in software development in 2017. He later taught foundational AI classes at Politehnica University of Timișoara and researched three-dimensional object detection and tracking for autonomous driving at Continental. His account of becoming an AI engineer traces his transition into computer vision, machine-learning infrastructure, and production AI, including roles at Everseen, CoreAI, and Metaphysic. More recently, he worked as a founding AI engineer at a San Francisco startup developing agents for financial services.

He founded Decoding AI Magazine in 2023 to teach the complete path from AI prototype to deployed, monitored software. The LLM Engineer’s Handbook, published in 2024 and coauthored with Maxime Labonne, develops that approach through an LLM Twin: a personalized application combining data pipelines, retrieval, fine-tuning, deployment, monitoring, and evaluation. His collaboration with Louis-François Bouchard subsequently expanded into agent-engineering education and open-source practical workshops.

• AI Research OS: Iusztin and Bouchard developed an open-source research-memory system that converts notes, saved articles, repositories, and other sources into persistent, project-specific context. Its file-based research wiki separates immutable source material from generated summaries, comparisons, conceptual notes, and a YAML index. Agents consult compact indexes and summaries before opening complete sources, reducing token consumption while keeping the system inspectable and independent of a vector database.

• Human-directed technical writing: His research-and-writing architecture combines human-specified topics, angles, audiences, reusable style profiles, and representative writing samples. Separate reviewer and editor contexts check drafts against source research and authorial intent; structured feedback prioritizes the writer’s instructions over stylistic preferences. Local agent skills can handle individual workflows, while Model Context Protocol servers simplify distributing shared tools, dependencies, and credentials.

• Calibrated agent evaluation: Iusztin treats an LLM judge as a binary classifier requiring real application outputs, human pass-or-fail labels, development and test splits, and precision, recall, and F1 measurements. Comparing development performance with held-out and live examples exposes overfitting; production traces covering tool calls, latency, cost, and outputs support debugging and regression detection.