ML Engineer, Frontier AI

About Us:

Here at Ambience, we never set out to be just another scribe. We’re building the AI intelligence platform that restores humanity to healthcare and drives meaningful ROI for health systems across the country.

Our technology helps providers focus on delivering great care by removing the administrative burden that pulls them away from patients and away from their most impactful work. Ambience delivers real-time coding-aware documentation and clinical workflow support across ambulatory, emergency and inpatient settings at the top health systems in North America.

Our teams operate relentlessly with extreme ownership to build the best solutions for our health system partners. We value candor, positivity and deep thought — and we expect a lot from each other because we know the problems we’re solving truly matter.

Ambience was ranked #1 for Improving the Clinician Experience in the KLAS Research Emerging Solutions Top 20 Report, recognized by Fast Company as one of the Next Big Things in Tech, named one of the best AI companies in healthcare by Inc., and selected as a Linked In Top Startup in 2024 and 2025. We’re backed by Oak HC/FT, Andreessen Horowitz (a16z), OpenAI Startup Fund, and Kleiner Perkins — and we’re just getting started.

As a Staff ML Engineer on the Frontier AI team at Ambience, you'll own the hardest model quality problems across our clinical AI products — foundational coding models, adaptive scribing, voice agents, long-context chart understanding, and clinical reasoning. This isn't a platform or infrastructure role. You'll set research direction, design learning loops, and drive end-to-end model quality improvements that compound over time.

Ambience ships advanced clinical AI in real-world healthcare settings. The models that power our products operate under constraints you won't find in typical ML roles — proprietary ontologies, messy EHR data, high compliance stakes, and clinician workflows where latency and accuracy both matter. You'll bring deep research instincts and engineering discipline to push the frontier on all of it.

Our engineering roles are hybrid - working onsite at our San Francisco office three days per week.

• Own foundational model research. Identify failure modes, form hypotheses, and drive architecture decisions on hard clinical AI problems — medical coding, adaptive scribing, chart understanding, and more.

• Build compounding learning loops. Design systems that turn real-world signals — clinician edits, coder corrections, audit outcomes — into fast, safe model improvements.

• Improve Chart Chat quality. Drive better grounding, smarter retrieval, and reasoning that holds up under the real diversity of clinical questions over complex longitudinal patient records.

• Push latency, accuracy, and cost simultaneously. Apply the right optimization levers — capability routing, distillation, speculative decoding, quantization — and know when each is safe.

• Contribute to population-level clinical reasoning. Help build toward a layer of clinical intelligence that reasons not just over individual patients, but across patient populations at scale.

• Stay at the cutting edge. Distill insights from recent research — particularly in RL, deep learning, and clinical NLP — and drive experiments that keep Ambience at the frontier of clinical AI.

• Deep RL and Deep Learning Expertise
  • 5+ years of ML engineering or applied research experience, with a strong track record of shipping model improvements in production.
  • Deep expertise in reinforcement learning and deep learning, developed in industry or a research setting.
  • Publications at top-tier venues (NeurIPS, ICML, ICLR, ACL, EMNLP, etc.) are a strong plus.
• Research to Production
  • Comfortable spanning research and engineering — architecture decisions, training runs, fine-tuning pipelines, and production deployment.
  • Experience with preference learning, RLHF, retrieval-augmented generation, or multi-label classification.
  • Strong Python fundamentals and experience with deep learning frameworks (PyTorch preferred).
• End-to-End Ownership
  • Can point to model quality improvements driven end to end: from identifying a failure mode…