Pocket
Research Engineer - Voice Intelligence
New
200000 –300000 USD / yearUSASeniormachine learningsignal processingtranscriptionspeaker identificationreal-time systems
Research-minded engineer to advance voice intelligence from capture to real-time experiences, focusing on improving voice stack and product capabilities.
Requirements
- Have 6–10+ years of experience shipping production systems with strong technical ownership.
- Are strong in applied ML/research engineering and can turn prototypes into robust product.
- Understand audio/voice fundamentals (signal processing basics helpful) and modern model/eval workflows.
- Care deeply about performance and reliability.
Nice to have
- Experience with streaming audio pipelines and real-time systems.
- Experience with LLM post-processing, structured extraction, and evaluation.
- Experience building tooling for labeling, dataset curation, and QA.
Conditions
- Work directly with us and learn fast
- Direct impact on how the company operates day to day
- High-trust, high-responsibility environment
- Competitive compensation
Other
- Pocket is building a category-defining platform for ambient intelligence. Voice is the highest-signal interface we have — and we’re looking for a research-minded engineer to push what’s possible with voice end-to-end: capture → understanding → real-time experiences → evaluation.
- Invent and ship improvements to our voice stack: diarization, VAD, noise robustness, transcription quality, speaker ID, and post-processing.
- Build new voice-powered product capabilities: better memory, better meeting/voice summaries, better action extraction, better personalization.
- Run tight research loops: define metrics, build eval sets, iterate on models/algorithms, and productionize what works.
- Improve real-time performance and reliability: streaming pipelines, latency budgets, fallbacks, and graceful degradation.
- Partner closely with product + design to translate voice capabilities into features people feel immediately.
- Step-function improvements in voice quality and perceived intelligence (measured and felt).
- A clear evaluation harness (offline + online) that prevents regressions and accelerates iteration.
- Voice features that ship reliably at scale with predictable latency.