AI Resident

The problem space

The loop we care about: production traces become data, data becomes training and evaluation, and better agents produce better traces. Projects live somewhere on that loop.

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.

What we're looking for

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.

Nice to have

Hands-on post-training with open models (TRL, veRL, OpenRLHF, or your own loop). Bonus points if you've debugged a reward-hacked run.

Built or trained in interactive agent environments (SWE, web, or tool-use gyms).

Large-scale trace analysis, data curation, or synthetic data work.

Serving and efficiency experience: vLLM/SGLang, distillation, quantization.

Multi-node GPU training, or the infra fluency to get there fast.

Publications, open-source work, or writing that shows how you think.

Security instincts: prompt injection, data governance, why a self-improving agent needs a fence.

Logistics

SF Bay Area, on-site/hybrid, half/full-time for the term. Flexible start.

Salary: $4,000 month

Compensation offered will be determined by factors such as location, level, job-related knowledge, skills, and experience. Certain roles may be eligible for variable compensation, equity, and benefits.

Ema Unlimited is an equal opportunity employer and is committed to providing equal employment opportunities to all employees and applicants for employment without regard to race, color, religion, sex, national origin, age, disability, sexual orientation, gender identity, or genetics.