Job Description

About the Team

You will join the TikTok USDS JV - Data Governance and Assurance Engineering team, the architects of the foundational infrastructure managing the end-to-end AI lifecycle. Our mission is to ensure the integrity, security, and compliance of our AI assets through automated frameworks and advanced content assurance standards.

Our team operates at the high-stakes intersection of TikTok's State-of-the-Art (SOTA) Recommendation Systems, Adversarial ML, and AI Governance. We are building the "safety operating system" for TikTok’s US data, replacing manual oversight with scalable, automated code that enforces safety and alignment across one of the world's most sophisticated algorithmic ecosystems.

About the Role

As an Machine Learning Engineer - AI Safety, you will be responsible for the technical "guardrails" and integrity of TikTok’s recommendation and General Purpose AI (GPAI) systems. You will build the high-performance engines that audit, evaluate, and mitigate risks in real-time.

Core Focus Areas:

- Recommendation & GPAI Assurance: Architecting auditable recommendation pipelines to ensure content delivery systems meet strict USDS compliance and safety requirements.

- Algorithmic Interpretability: Enhancing the transparency of recommendation signals to ensure model decisions are explainable and aligned with safety protocols.

- Continuous Monitoring & Mitigation: Developing automated pipelines to evaluate, monitor, and fine-tune GPAI/LLM deployments to proactively mitigate risks such as bias, toxicity, and misinformation.

Responsibilities

- Automated Safety Evaluation: Design and scale high-throughput evaluation engines to benchmark LLM and Recommendation Model outputs against safety, bias, and accuracy protocols.

- Adversarial Robustness: Build and maintain continuous monitoring frameworks to proactively identify vulnerabilities in recommendation logic and model behavior.

- Governance-as-Code: Architect metadata frameworks and control planes for ML pipelines, ensuring training data and model weights adhere to purpose-limitation and safety constraints.

- Real-time Mitigation: Develop low-latency safety filters and "circuit breakers" that can intervene in real-time recommendation streams (ms-level latency) without compromising system performance.

- Observability & Lineage: Create centralized platforms that provide real-time observability into model decisions, ensuring every algorithmic "promotion" is traceable, auditable, and compliant.