Principal Product Manager

The base model gets you capability. Post-training is what turns that capability into a model people actually want to use; one that follows instructions well, has the right tone and judgment for Copilot's surfaces, refuses the right things, and keeps getting better release over release.

We're looking for a Principal Product Manager to sit inside that process. You'll be embedded directly with applied researchers and ML engineers running RLHF, fine-tuning, and preference-tuning pipelines, reading model outputs, calling out what's wrong, and turning that into concrete training priorities. You're the person who brings the end user's voice into a room that's otherwise all metrics and loss curves.

This is a technical, hands-on PM role, not just a roadmap-and-requirements role. You'll spend real time in transcripts and eval dashboards, and you'll need enough grounding in how post-training actually works (RLHF, DPO/preference optimization, instruction tuning, data curation) to have a credible point of view with the research team.

Microsoft's mission is to empower every person and every organization on the planet to achieve more. As employees we come together with a growth mindset, innovate to empower others, and collaborate to realize our shared goals. Each day we build on our values of respect, integrity, and accountability to create a culture of inclusion where everyone can thrive at work and beyond.

• Embed in the post-training loop: work day-to-day with research and applied ML teams during fine-tuning and RLHF cycles, reviewing model outputs and giving structured feedback on quality, tone, and behavior

• Define what "good" means for Copilot's models: set the priorities for capability, personality, and safety tradeoffs across surfaces (Microsoft 365 Copilot, Copilot Studio agents, Windows) - where "good" is often genuinely ambiguous and hasn't been decided before

• Run RLHF/preference data campaigns: partner with data and labeling teams to scope what human preference data gets collected, for which behaviors, and why

• Build and maintain behavior evals: translate qualitative judgment calls ("this response felt too hedgy," "this refused when it shouldn't have") into evaluation sets that can be tracked release over release

• Own the feedback loop: turn user research, enterprise customer feedback, and production incident learnings into concrete post-training priorities - closing the gap between "users are unhappy with X" and "the next fine-tune addresses X"

• Make the tuning tradeoffs explicit: helpfulness vs. caution, consistency vs. personality, latency/cost vs. quality - and drive alignment across research, safety, and product leadership on where the line sits

• Represent Responsible AI and enterprise requirements into training priorities - compliance behaviors, refusal policies, and brand voice all have to be reflected in what the model is actually tuned to do.