About the role
About Ema Ema is building the world’s leading Agentic AI platform to transform enterprise productivity. We enable organizations to delegate repetitive tasks to Ema, the Universal AI Employee, delivering 10x gains in workforce efficiency, across functions. Founded by former executives from Google, Coinbase, Flipkart, and Okta, our team includes engineers from premier tech companies and graduates of Stanford, MIT, UC Berkeley, CMU, and IITs. We are backed by industry leading investors including Accel, Naspers/Prosus, Section32, and angels like Sheryl Sandberg and Dustin Moskovitz. Headquartered in Silicon Valley and with offices in London, Bangalore and Vancouver and Bangalore, Ema is at the frontier of what Agentic AI can do in production — we ship real systems that run real business processes at scale. The residency 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. 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.
