Machine Learning Engineer - Agentic AI

DiDi Labs•San Jose, CA

9h•$169,783 - $282,264

About The Position

DiDi Autonomous Driving is seeking a Machine Learning Engineer to develop agentic AI approaches for improving autonomous driving software. The role explores AI agents that can propose changes to policy code, data, training recipes, and model code; evaluate results; diagnose failures; and turn verified fixes into reusable skills and data. Using DiDi Voyager’s real-world driving data and large-scale evaluation infrastructure, the goal is to create an improvement loop that generalizes across scenarios, passes safety gates, and delivers improvements to vehicles. The work combines machine learning, agentic systems, and autonomous driving, and contributes to both onboard architecture and the development process for driving software.

Requirements

Strong background in machine learning, such as deep learning, reinforcement learning, or decision-making systems.
Hands-on experience building LLM agent systems, such as tool use, multi-agent orchestration, or automated evaluation loops.
Experience training or fine-tuning LLMs for agentic workflows, including trajectory data curation, supervised fine-tuning, or reinforcement learning from execution feedback.
Strong Python skills and working knowledge of C++.
Rigorous experimental habits, including controlled comparisons, held-out evaluation, and a clear distinction between measured results and assumptions.
Ability to own a system from research prototype through production.

Nice To Haves

Experience building an agentic harness or self-improving loop for a specific product or domain.
Experience in autonomous driving or robotics, especially planning, prediction, simulation, or sim-to-real transfer.
Familiarity with code-as-policy, program synthesis, LLM-driven optimization, or evolutionary search.
Publications at top machine learning or robotics venues such as NeurIPS, ICML, ICLR, CVPR, RSS, or CoRL.
BS, MS, or PhD in Computer Science, Machine Learning, Robotics, or a related field.

Responsibilities

Build agentic harnesses that give agents permissioned access to the driving stack, data, training infrastructure, and evaluation environments, with automated scenario setup and pass/fail checks.
Develop recursive self-improvement loops involving failure mining, candidate generation, evolutionary search, and targeted data generation focused on policy weaknesses.
Curate a reusable library of skills from verified fixes and check that each skill is beneficial before promotion.
Create training data from agent development trajectories and fine-tune LLMs to work with the software stack, tools, and codebase.
Advance agent-as-policy approaches, including writing policy logic, curating data, adjusting training recipes and objectives, and tracing driving failures to their causes.
Shape the onboard architecture by determining what runs on the vehicle versus offline, accounting for latency, compute, determinism, and safety constraints.
Collaborate with prediction, planning, simulation, and onboard teams to turn prototypes into production workflows.