Research Engineer - Voice Intelligence

Description

About Pocket

At Pocket, we’re building the future of human memory.

Some of the most valuable information in our lives disappears when a conversation ends: an idea, a decision, a detail we wish we could remember. Pocket helps us hold onto it.

We’ve built an AI-powered device that captures real-world conversations and turns them into searchable transcripts, summaries and next steps. Backed by Accel and Y Combinator (W26), we’ve already sold 500,000 devices.

Our ambition goes beyond recording. We believe the things you learn, discuss and experience can become useful context for the AI tools you already use, and for products that don’t exist yet.

About the Role

Real conversations are messy. People interrupt, rooms are noisy, speakers sound similar and a missed detail can change the meaning of a summary.

You’ll improve the voice systems that make Pocket useful in those conditions. The work connects applied research to production: define the problem, build an evaluation, test an approach and ship what works. You’ll partner with product and design so technical progress becomes something customers can notice.

What You’ll Own

Improvements to diarization, voice activity detection, noise robustness, transcription, speaker identification and post-processing.

Voice-powered capabilities such as better summaries, action extraction, memory and personalization.

Research loops with explicit metrics, representative evaluation sets and repeatable comparisons.

Moving promising models and algorithms from prototypes into reliable product features.

Streaming performance, latency, fallbacks and behavior when systems cannot produce a confident answer.

Requirements

• 6–10+ years shipping production systems with strong technical ownership.
• Applied ML or research-engineering experience, including turning prototypes into dependable software.
• A working understanding of audio and voice fundamentals, modern models and evaluation workflows.
• Care for performance, reliability and the failure cases that a demo can hide.
• The ability to work with product and design and explain technical tradeoffs clearly.
• Bonus Points
• Streaming audio and real-time systems; LLM post-processing and structured extraction; or tools for labeling, dataset curation and quality assurance. Signal-processing fundamentals are useful too.
• What Success Looks Like
• Voice quality improves on meaningful evaluations and in everyday use. An offline and online evaluation system catches regressions. New capabilities reach customers with predictable latency and reliable behavior.

Benefits

• You’ll work on voice captured by a real consumer device, with a direct path from an experiment to a shipped feature. The job gives you room to investigate hard problems without losing the connection to the person relying on the result.