Machine Learning Engineer 5

Adobe•San Jose, CA

1d•$211,800 - $306,625•Onsite

About The Position

Join us at Adobe as a Machine Learning Engineer (MLE 50) on the Adobe Brand Intelligence Predict team in San Jose, CA! Help us build the next generation of synthetic audiences, LLM-powered simulated consumers that let the world's biggest brands pre-test ads, campaigns, and content before a single dollar is spent.

Requirements

Substantial hands-on experience building LLM-based applications in production.
Demonstrated experience designing and shipping complex inference harnesses on top of large language models (agentic systems, structured reasoning, sampling/decoding strategies, RAG).
Hands-on experience fine-tuning LLMs with techniques including SFT, preference optimization (DPO/GRPO) and modern post-training tradeoffs.
Experience with RLHF, RLAIF, or RL-based state alignment of LLMs.
Proven track record of building evaluation datasets and harnesses — you have opinions about what makes an eval load-bearing versus theater.
Proficiency in Python and strong grounding in data structures, algorithms, and modern ML tooling (PyTorch, Hugging Face, vLLM, W&B or equivalents).
Hands-on knowledge of MLOps practices and pipelines.
Familiarity with cloud ML services (AWS, GCP, Azure).
Shipped a customer-facing Gen AI feature from proof-of-concept to production end-to-end.
MS or PhD or equivalent experience in Computer Science, Machine Learning, or a related technical field, or equivalent experience.

Nice To Haves

Prior work on synthetic audiences, persona simulation, or LLM-based human behavior modeling.
Familiarity with the synthetic audiences research literature (e.g., silicon samples, generative agents, SubPOP, HumanLM, DeepBind).
Experience with public opinion or survey data (GSS, ANES, WVS, MIDUS) or panel-based consumer research data.

Responsibilities

Design, build, and ship LLM-powered systems that simulate consumer audiences end-to-end, from proof-of-concept to production.
Develop complex inference and reasoning harnesses on top of frontier LLMs, agentic flows, persona conditioning, retrieval, and sampling strategies tuned for distributional fidelity.
Fine-tune LLMs on survey, panel, and behavioral data to improve alignment with real-world audience distributions; own the full loop from data curation through eval.
Build the evaluation datasets, benchmarks, and harnesses that define what "good" means for synthetic audience quality - distributional fidelity, behavioral validity, subgroup calibration.
Partner with product management, applied science, and engineering to translate a fast-moving research literature into shipping product features.