AI Resident, Ema

• No specific degree required. Strong undergrads, grad students, and self-taught builders are all welcome; what matters is demonstrated depth in ML or agent systems
• Solid ML fundamentals and strong engineering: Python, PyTorch, and the discipline to ship in a large production codebase
• Real depth in at least one of: post-training (SFT/DPO/GRPO-family RL), reward modeling or LLM judges, agent and tool-use systems, retrieval and memory, eval design. One area you can teach us beats five you've touched
• Statistical literacy. You can size an experiment, and you know 25 samples at one seed is a datapoint, not a result
• Honest measurement as a habit. You'd rather kill your own feature with a clean experiment than ship it on a hunch

Desirable

• You own one hard problem end to end. You write the proposal, build the system, design the evaluation, ship behind a gate, and finish with a write-up of what turned out to be true, including the parts that didn't work. You'll sit in the production codebase with a senior mentor and real production data. Recent residents have shipped self-improving harnesses, inference-cost work, agent memory, and eval infrastructure. Your project gets scoped with you, not handed to you
• Harness and inference-time work. Context engineering, tool and skill design, orchestration, and deciding where extra inference compute actually pays. Self-improvement loops run behind hard fences
• Post-training for agents. SFT on curated trajectories, preference optimization, RL on real agent tasks. Reward design where outcomes are verifiable, process vs. outcome supervision, distilling frontier behavior into cheaper models
• Environments and rewards. Turning enterprise workflows into training and eval environments: fixture tenants, simulated users (some of whom get impatient and leave), verifiable rewards, and defenses against reward hacking. Agents will exploit a lazy grader
• Data engines. Mining production agent-steps into training and eval corpora: failure mining, labeling with calibrated judges, synthetic augmentation that stays useful
• Evaluation. Behavior-level benchmarks from real workflows, LLM judges calibrated against human labels, reliability statistics for stochastic agents
• Efficiency. Routing, ensembles, caching, small-model specialization. Quality per dollar is a research metric here