The hard problems are the ones frontier models don't solve on their own: maintaining an accurate picture of a learner over weeks and months, deciding what to teach next and when to hold back, keeping long-running conversations useful rather than merely pleasant, and verifying that generated teaching is correct before a learner ever sees it. 3+ years as a software engineer, with substantial hands-on experience building with LLM APIs (Claude, OpenAI, or similar): agentic workflows, tool use, structured output, long-context and memory patterns.