Research Engineer, Privacy and Anonymization at hud in San Francisco

Job description

• Application
• About HUD
• HUD is building infrastructure to create RL training data and evals for frontier AI agents, as well as a marketplace to sell these to frontier labs through the HUD marketplace. Our platform is used by frontier labs, Fortune 500 companies, and startups. We’ve raised $16M from top VCs and were YC W25.
• We’re looking for a Research Engineer to build the privacy and anonymization systems that make sensitive, real-world data safe and useful for AI training. You’ll develop methods to detect and remove PII, secrets, and other sensitive information from raw data before it enters our processing and synthetic data pipelines. You’ll own the full pipeline for protecting privacy without destroying the structure and signal that make data valuable for training agents.

Responsibilities

• Build systems to detect PII, quasi-identifiers, credentials, and other sensitive information and design transformations based on the data type and downstream use case
• Develop and benchmark detection approaches that combine rules, statistical models, classifiers, and LLM-based methods
• Build production pipelines that anonymize raw data before it enters downstream processing, training, evaluation, or synthetic data generation workflows
• Create evaluation frameworks that measure privacy risk and retained data utility, including recall-weighted metrics, leakage tests, and adversarial re-identification attempts
• Design systems that remain robust to new data sources, schema drift, unusual formats, and sensitive information embedded in unexpected fields
• Work with engineering, research, operations, and customers to translate privacy requirements into practical technical policies and safeguards

Requirements

• Strong proficiency in Python and experience building reliable production data or ML systems
• Experience with information extraction, named-entity recognition, classification, or related methods for detecting rare or sensitive content
• Strong experimental instincts and the ability to compare approaches across recall, precision, latency, cost, and downstream data utility
• An understanding of the difference between redaction, masking, pseudonymization, anonymization, and synthetic data—and when each is appropriate
• High attention to detail and the ability to reason about subtle leakage paths, edge cases, and adversarial failure modes
• Built data processing pipelines end-to-end without a fully prescribed roadmap
• Hands-on experience with privacy-enhancing technologies such as differential privacy, k-anonymity, secure aggregation, format-preserving encryption, etc.
• Worked with sensitive data in areas such as healthcare, finance, or security
• Built low-latency or high-throughput ML inference and data-processing systems
• Worked in unstructured problem spaces and take ownership from early research through production deployment
• Early-stage startup experience and strong communication skills for collaboration across teams and time zones
• We prioritize technical aptitude and learning potential over years of experience. Motivated candidates are encouraged to apply even if they don't meet all criteria.

Skills

Required

• Python
• Production data or ML systems
• Information extraction
• Named-entity recognition
• Classification
• Experimental design
• Data anonymization techniques
• Attention to detail

Preferred

• Differential privacy
• k-anonymity
• Secure aggregation
• Format-preserving encryption
• Data-processing systems
• Early-stage startup experience
• Communication skills

Benefits

• Competitive compensation
• 100% covered top-of-the-line medical, dental, and vision from Blue Shield of CA (US employees)
• Lunch and dinner when you’re in the office (in-office employees)
• Company-wide holiday break (Christmas Eve to New Year’s Day) on top of PTO and paid holidays
• Other perks including an Equinox membership, 401k, and commuter benefits (US employees)
• Unlimited* access to tokens for ChatGPT, Claude Code, Cursor, etc. *By unlimited, we mean no one on our token usage leaderboard has ever hit a limit. So we have no idea what the limit is.