Machine Learning Engineer, Safety and Customer Care AI

Written by the T&S Network from Lyft's posting.

Builds and post-trains open-source models and multi-step AI agents for Lyft's Safety and Customer Care team, which handles rider and driver support and safety case management. The work covers the evaluation and data flywheel that proves these systems are safe to run in a customer-facing, safety-critical setting. The team serves as Lyft's primary direct touchpoint with riders and drivers, with over 1.7 million monthly human and AI interactions.

What you'd do

• Adapt open-source LLMs for SCC tasks through supervised fine-tuning, LoRA, and preference tuning.
• Design and build support agents and end-to-end safety case management agents in LangGraph or equivalent frameworks.
• Assemble human and synthetic datasets and refine model output quality on live support and safety work.
• Build evaluation frameworks, golden sets, rubric grading, LLM-as-judge, and regression tests.
• Ship models and agents into real-time production with monitoring and guardrails.
• Convert feedback from customer interactions into training data and learning signals.
• Apply traditional ML like classification, ranking, and gradient-boosted trees where appropriate, with product, ops, and data science.

What it asks for

• At least 3 years in applied ML/AI, plus an MS or PhD in CS, ML, AI, or a related technical field.
• Post-training experience with open-source models, including SFT, LoRA, RLHF, RLAIF, and RLVR.
• Have built and launched agents on LangGraph or a comparable framework.
• Designed metrics and built offline and online evaluation for generative systems.
• Have put ML/AI systems into real-time production.
• Strong Python and hands-on PyTorch.
• Applied ML/AI to customer support or trust and safety work, like agent assist, routing, or abuse detection (preferred).

Worth knowing

• Role works with a safety-critical customer-facing system serving riders and drivers.

Terms in this role

RLHF.

Reinforcement learning from human feedback: training an AI model on people's ratings of its answers.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Lyftnorth_east before you apply. If we got something wrong, report this listing.