Research Engineer, Post-Training

present of AI; United States; Remote; USD 120,000 - 180,000; Full time

Job summary

Deepgram is seeking a research engineer focused on post-training to improve base speech and language models through fine-tuning, preference tuning and rigorous evaluation. You will design experiments, measure outcomes and ensure the model behaves well on real tasks for developers.

The role emphasizes hands-on training with Python and PyTorch, careful experiment tracking and clear reporting to keep models useful and reliable in production voice applications.

Qualifications

• Hands-on model training experience with fine-tuning and evaluation.
• Strong Python and deep learning skills.
• Experience with experimenting and measurement design.

Responsibilities

• Develop post-training workflows for speech and language models.
• Evaluate model outputs and identify failure modes.

Skills

Python

PyTorch

Deep learning

Experiment design

Job description

A research engineer in post training usually takes a base speech or language model and improves it through fine tuning, preference tuning and evaluation, so the model behaves well on real tasks. A strong candidate brings hands on model training experience, solid Python and deep learning skills, and careful experiment design with clear measurement. At a company that builds its own voice models, this work directly shapes how well transcription and voice agents perform for developers.

How it fits the AI era

This role sits at the center of the AI era, since post training is where raw models become useful products. Much routine training and evaluation work is increasingly automated, but judging model quality, spotting failure modes and deciding what to optimize stay human. The engineer spends more time on data quality, evaluation design and tradeoffs than on writing boilerplate.

How it builds AI experience

The work builds direct experience with fine tuning, preference optimization and evaluation of speech and language models, which are core skills for applied AI roles. Building small post training experiments on open models and measuring results is a practical way to prepare. Experience with speech data and model behavior on audio tasks adds a useful edge for a voice focused company.

Worth showing

• Python and PyTorch
• Speech or language models
• Careful experiment tracking

This is our own read, written from the employer's posting and the company's Spotlight. The employer's posting is the source of truth for the role.

About Deepgram

A speech AI company that builds its own voice models from scratch, so developers can add transcription and voice agents to their products by API.