Design GenAI Fine-Tuning and Agent Tradeoffs

You are interviewing for a software engineering role involving generative AI infrastructure and quantitative applications. The interviewer wants to understand how you make practical production tradeoffs, not just whether you have used large language models. Answer the following as a system and machine learning design discussion: 1. When would you choose full-precision supervised fine-tuning, LoRA, or QLoRA? 2. How do data size, model size, GPU memory, training budget, latency requirements, and target quality affect that choice? 3. If the system had to scale to larger models, more data, lower latency, or tighter cost constraints, what would you change? 4. What is the role of an agent framework in production? Discuss structured outputs, schema validation, tool calling, orchestration, state management, evaluation, observability, and failure handling. 5. How would you prevent an agent from going out of control, silently failing, producing invalid outputs, or making unsafe tool calls? Ground your answer in concrete engineering decisions, metrics, and tradeoffs.