Senior Applied Scientist - Agentic AI

At Oracle Analytics, we are building the next generation of enterprise AI products to enable intelligent data analysis eraging our foundational strengths in data management and enterprise software applications, we are advancing our platforms and applications by deeply embedding cutting‑edge agentic AI, generative AI, and innovations in machine learning and optimization. We are seeking a Senior Applied Scientist to perform innovation in learning from human feedback (LFHF) and user preference modeling, with a strong focus on in‑context learning and post training for large language models.

You will design data and feedback strategies, build preference/reward models, and develop post training pipelines (e.g., SFT, DPO, RLHF/RLAIF) that deliver safe, high‑quality, and cost‑efficient enterprise AI experiences. You will partner closely with research engineers and product teams to ship aligned models to production, instrument rigorous evaluation, and drive measurable customer and business impact.

• Perform end to end LFHF programs: define annotation rubrics, sampling strategies, and quality controls; design rater guidelines and human in the loop workflows in collaboration with product/UX and data engineering.
• Build preference and reward models: pairwise and listwise modeling, win rate optimization, uncertainty estimation, and active learning to improve sample efficiency and data quality.
• Develop post training pipelines: supervised fine tuning (SFT), direct preference optimization (DPO/IPO/ORPO), RLHF/RLAIF, and distillation—balancing quality, safety, latency, and cost for enterprise workloads.
• Advance in‑context learning: retrieval augmented prompting, dynamic few shot selection, tool use/orchestration aware prompting, instruction following, and mitigation of ICL brittleness and context overflow.
• Optimize inference and efficiency: PEFT/LoRA/QLoRA, quantization, speculative decoding, caching, and distillation for scalable deployment on Oracle infrastructure.
• Evaluate rigorously: establish offline/online metrics, pairwise and rubric based human evals, red teaming, safety/guardrail tests, A/B experiments, and win rate tracking; perform offline policy evaluation where applicable.
• Ensure safety, privacy, and compliance: apply content safety policies, guardrail configuration, PII handling/redaction, differential logging, and model governance appropriate for regulated enterprise settings.
• Productionize solutions: collaborate with platform teams to ship models and evaluation services; implement observability, telemetry, canarying, rollback, and lifecycle management.
• Stay current with research and translate advances into production differentiators; mentor teammates and contribute to a culture of scientific rigor and impact.

• MS, PhD (preferred) in Computer Science, Machine Learning, Statistics, Electrical Engineering, or related field with a focus relevant to LFHF, reinforcement learning, NLP, or human AI interaction.
• Experience (industry or applied research) building and deploying ML systems, including LLM post training and evaluation.
• Demonstrated expertise in learning from human or AI feedback: data/rubric design, preference/reward modeling, and optimization methods (e.g., SFT, DPO, RLHF/RLAIF).
• Strong background in in context learning, prompt/program design, retrieval augmented generation, and model alignment for accuracy, safety, and robustness.
• Proficient in Python and modern ML stacks:
  PyTorch/JAX, Transformers, and libraries for post training and evaluation; solid software engineering practices and experimentation discipline.
• Track record publications in top venues (e.g., NeurIPS, ICML, ICLR, ACL, EMNLP, NAACL).

• Experience designing at scale data pipelines for feedback collection, active learning, and rater operations; familiarity with label quality auditing and bias/variance trade offs.
• Knowledge of bandits/off policy evaluation, causal inference for policy changes, and statistical testing for online experiments.
• Familiarity with LLM efficiency and serving: tensor/graph optimization, KV cache management, batching strategies, and throughput/latency trade…