Product Manager - AI Evaluation & Agentic Platform

Meta; Menlo Park (CA); Hybride; USD 180 000 - 240 000; Plein temps

Résumé du poste

Meta’s ATA Product team is building Metamate, an internal platform for improving agentic AI across the company. This role focuses on measuring and raising the quality of agentic systems, defining what good means for agents, and owning the eval verdict that gates model upgrades and launches.

The position is early-stage, high-leverage, requiring collaboration with engineering, data science, ML, and partners across ATA and MSL to turn real usage signals into a continuous improvement loop that

Qualifications

• 5+ years in industry with at least 2 in product management.
• Experience building evals for AI products end to end: task sets, rubrics, instrumentation.
• Strong quantitative skills and ability to define metrics where none exist.
• Comfort with ambiguous problem spaces and rapid learning.
• Experience with RLHF or human-in-the-loop systems is a plus.

Responsabilités

• Define metrics and the eval verdicts that gate model upgrades and launches.
• Collaborate with data science and ML engineering teams across ATA and MSL.
• Own end-to-end evaluation development: rubric design, instrumentation, analysis.
• Drive measurable impact through continuous improvement loops sprint to sprint.

Connaissances

Quantitative analysis

Product management

Experimentation design

AI/ML familiarity

Cross-functional collaboration

Eval metrics

Formation

Bachelor's degree in STEM or related field

Outils

Experimentation platforms

A/B testing

Description du poste

Meta’s ATA Product team is building Metamate, an internal platform for improving agentic AI across the company. This role focuses on measuring and raising the quality of agentic systems, defining what good means for agents, and owning the eval verdict that gates model upgrades and launches.

The position is early-stage, high-leverage, requiring collaboration with engineering, data science, ML, and partners across ATA and MSL to turn real usage signals into a continuous improvement loop that