About The Role

ABOUT THE ROLE

You'll own the transcription pipeline end-to-end at an early-stage ambient intelligence consumer startup — a cloud-based ASR system with a narrowly scoped on-device component. As one of the company's first US engineering hires, you'll work hands-on with senior product and business leadership to build, tune, debug, and ship pipeline improvements directly.

WHAT YOU'LL DO

Build and iterate on the cloud-based ASR pipeline, from audio capture through post-processing, in production at scale.
Own ASR quality and reliability end-to-end, shipping measurable improvements across latency, small-word accuracy, and voice-print reliability.
Work across data, training/fine-tuning, evaluation, and deployment to translate product feedback into shipped pipeline changes.
Collaborate closely with overseas R&D, hardware, and supply-chain teams across time zones.
Partner with a product-focused backend engineer on shared pipeline surfaces.
Operate with minimal specification, turning lightweight asks into concrete, production-ready improvements.

WHAT WE'RE LOOKING FOR

3+ years building and tuning transcription/ASR pipelines end-to-end in production, primarily in cloud-based settings.
Demonstrated ownership of production ASR systems through the full lifecycle: data preparation, model training/fine-tuning, evaluation, and deployment.
Hands-on experience with latency-sensitive or streaming audio/ASR pipelines.
Track record of shipping pipeline improvements from design through deployment based on real production usage data.
Comfort debugging transcription quality issues (small-word accuracy, voice-print reliability, latency) in live systems.
Experience in early-stage or founding engineering environments — ships without large team support or fully-specified requirements.
On-device or embedded ML experience (Core ML, TensorFlow Lite, or similar frameworks).
Prior experience building wearable, hardware, or robotics device products.
Background at an AI-native consumer application focused on transcription or audio.
Experience building agent or LLM-based product features (tool use, memory, retrieval systems).
Cares how transcription feels to use — not just how it benchmarks — and makes latency/accuracy tradeoffs independently.