Staff AI Scientist, Fiddler AI

• 7+ years of applied AI experience, with a strong track record of taking models from research to production
• Experience in LLM or Agentic Evals, Guardrailing
• Deep expertise training and fine-tuning classifier models, including modern encoder architectures (BERT-family, ModernBERT, etc.) and LLM-as-classifier approaches; clear understanding of the tradeoffs between them
• Hands-on experience with dataset development as a first-class engineering discipline: sourcing, labeling, synthetic generation, adversarial augmentation, and quality control
• Strong applied experience with LLMs and agentic systems – prompting, fine-tuning, and evaluation
• Proficiency in Python and the modern ML stack (PyTorch, Hugging Face, common training/serving frameworks)
• Comfortable working in production environments and partnering with backend and platform engineers on real-time inference, monitoring, and rollout
• Excited by the prospect of using AI coding tools in your own workflow to push the limits of what one engineer can ship – responsibly, with a clear eye on quality, security, and the failure modes these tools introduce
• Excellent written and verbal communication; able to explain research tradeoffs to engineers, PMs, and customers
• Ability to work in our Palo Alto office 2-3 days a week

Desirable

• You’ll join a tight-knit, highly collaborative AI Science team that partners closely with some of the world’s leading enterprise AI organizations
• We operate in a hybrid capacity, but stay deeply connected through constant communication, collaboration, and shared purpose
• Our team thrives on knowledge sharing, peer learning, and collective problem solving – no one works in a silo
• We celebrate each other’s successes, support one another through complex challenges, and take pride in helping our customers achieve real-world impact with Fiddler’s AI Observability platform
• Every project is a team effort, and every win is shared
• If you love working alongside smart, driven peers who genuinely care about both customer success and each other’s growth, you’ll feel right at home here
• Lead applied research and development. Lead applied research and development for the models and datasets at the core of Fiddler's Trust Service and suite of guardrail classifiers and evaluators that customers depend on to keep their LLM and agentic applications safe, accurate, and compliant in production
• Partner closely with other engineering teams, Product, and Customer Success. You’ll build strong relationships with customer data science and ML engineering teams, supporting their AI observability journey and ensuring they realize measurable value from Fiddler
• Design, train, and ship production classifiers for safety, security, and quality detection (e.g., prompt injection, jailbreaks, PII, hallucination, faithfulness) under strict latency and cost constraints
• Lead the development of synthetic and adversarial dataset pipelines, including novel methods for generating, filtering, and validating data that exposes failure modes our models need to learn
• Drive the technical direction of generative insights – the LLM- and agent-powered analysis layer that helps customers diagnose what's going wrong in their AI applications and why
• Contribute to the evaluation and experimentation infrastructure that lets the AI Science team and our customers reliably measure model quality, regression, and drift across rapidly evolving model populations
• Explore reinforcement learning and preference-based methods where they offer real leverage over supervised baselines
• Collaborate with Backend and Platform engineers to take research prototypes from notebook to a hardened, scaled, observable service
• Partner with Product, Solutions Engineering, and Customer Success to translate enterprise customer needs into research problems and translate research results back into product
• Mentor AI Scientists on the team, raise the technical bar through code review and design review, and represent Fiddler externally through publications, talks, or open-source contributions when appropriate