Applied ML Engineer, Voice agents

Retrievant SAS; Boston (MA); Hybrid; USD 120,000 - 180,000; Full time

Benefits offered by this job

Equity

Health coverage

Office in Boston

Leave and learning budget

Relocation support

Job summary

Retrievant SAS is seeking an engineer to improve voice agents for difficult healthcare conversations. You will enhance ASR robustness, build hold/IVR/voicemail classifiers, and design state-machine conversation policies evaluated offline and online.

You will create an automated evaluation suite, own metrics, and ship updates weekly. The role involves listening to calls and iterating based on results, with options for remote or on-site collaboration.

Qualifications

• Experience with speech systems in production (ASR, TTS, or dialogue).
• Strong Python and PyTorch; comfortable with data pipelines and evals.
• You can read a transcript and know why the call failed.

Responsibilities

• Improve ASR robustness on real telephony audio; build the hold/IVR/voicemail classifiers.
• Design conversation policies as explicit state machines with LLM-generated turns, and evaluate them offline and online.
• Build the eval suite: transcripts, outcomes, specialist labels. Own the metrics.
• Ship weekly. Listen to the calls.

Skills

ASR systems

Python

PyTorch

Data pipelines

Evaluation metrics

Job description

Make our voice agents better at the hardest conversations in healthcare: hold music, IVRs, accents, interruptions and a receptionist having a bad day.

What you'll do

• Improve ASR robustness on real telephony audio; build the hold/IVR/voicemail classifiers.
• Design conversation policies as explicit state machines with LLM-generated turns, and evaluate them offline and online.
• Build the eval suite: transcripts, outcomes, specialist labels. Own the metrics.
• Ship weekly. Listen to the calls.

What we're looking for

• Experience with speech systems in production (ASR, TTS, or dialogue).
• Strong Python and PyTorch; comfortable with data pipelines and evals.
• You can read a transcript and know why the call failed.

How we hire

Four steps, two weeks if you want it that fast. A 30-minute call with the hiring manager. A practical exercise related to the role (paid, if it takes more than two hours). A half-day with the team, in person or remote, including a session listening to real retrieval calls. Then references and an offer.

• Meaningful equity at a seed-stage company
• Full health coverage for you and dependents
• Paris and Boston offices, or remote within EU / US time zones
• Six weeks' leave, plus local public holidays
• Annual learning budget, hardware of your choice, relocation support