GenAI Quant SWE Manager Deep Dive

Manager round for a GenAI / quant-infra SWE role, focused on practical LLM fine-tuning tradeoffs, production agent frameworks, prior scope, scale, latency, quality, cost, and reliability decisions. Requirements This round is a manager / technical deep dive rather than a coding prompt. Topics include: Role scope and how past experience maps to GenAI / quant-infrastructure work. When to choose full-precision SFT versus LoRA versus QLoRA. How data size, model size, GPU memory, latency, quality target, and cost constraints affect that choice. Agentic frameworks such as Pydantic-style structured output or ADK-style orchestration. Production concerns: schema validation, tool calling, orchestration, state management, evaluation, observability, avoiding uncontrolled agent behavior, and avoiding silent failure. Past project scope, concrete tech stack, why a particular solution was chosen, and how the solution would change at larger scale or tighter latency. Notes High-level "I used LLMs" answers are not enough. The round probes whether the candidate has made real tradeoffs under constraints. Be concrete about hardware, model size, training budget, quality target, and serving latency. Treat agent reliability as an engineering system: typed interfaces, validation, logs, evals, fallback behavior, and monitoring. Preparation Prepare one fine-tuning decision matrix covering full SFT, LoRA, and QLoRA. Write a short architecture for a production agent workflow: inputs, schema, tools, state, validation, evals, and observability. Rehearse a past GenAI project at three levels: business goal, technical stack, and scale/cost/quality tradeoffs.