ML Engineer — Speech

Build the real-time speech and language pipeline behind the coach — streaming transcription, speaker separation, and the models that detect objections and intent fast enough to fire a cue before the rep flinches. You care about sub-second latency as much as accuracy.

• Build and tune streaming ASR, diarization, and intent/objection-detection pipelines for live calls.
• Drive down end-to-end latency while holding accuracy; own the speed/quality trade-offs.
• Evaluate and integrate speech and LLM providers; build evals and guardrails for production quality.
• Instrument the pipeline so quality and latency regressions are caught before reps feel them.
• Partner with product to turn detected signals into cue cards that actually help mid-call.

• 4+ years in ML/applied ML, with production work in speech, NLP, or real-time inference.
• Strong Python and experience deploying models in low-latency, streaming settings.
• Practical grasp of ASR/diarization and modern LLM tooling and prompting.
• Rigorous about evaluation: you measure before you ship and after.
• Comfortable owning ambiguous, research-flavored problems end to end.
• Experience with streaming ASR providers (Deepgram, AssemblyAI) or self-hosted speech models.
• WebRTC/audio engineering or DSP background.
• Building LLM evals, judges, or RAG systems in production.

Our process has a few stages — it's a real role, so we take it seriously on both sides:

• Intro call with the hiring manager — the role, your background, and what you want next.
• Team conversations — meet the people you'd work with day to day.
• A practical exercise drawn from real work — no whiteboard puzzles, no weekend-eating take-homes.
• An on-site visit to our Saint Paul office to meet the team in person.
• Final conversation and references, then an offer.

We move as quickly as we can and keep you posted at every step.

If you'd be great at this work, tell us how you'd approach it. We read every note.