Model Evaluation QA Lead

Deepgram

About The Position

As Model Evaluation QA Lead, you’ll be the technical owner of model quality assurance across Deepgram’s AI pipeline—from pre-training data validation and provenance through post-deployment monitoring. Reporting to the QA Engineering Manager, you will partner directly with our Active Learning and Data Ops teams to build and operate the evaluation infrastructure that ensures every model Deepgram ships meets objective quality bars across languages, domains, and deployment contexts. This is a hands-on, high-impact role at the intersection of QA engineering and ML operations. You will design automated evaluation frameworks, integrate model quality gates into release pipelines, and drive industry-standard benchmarking—ensuring Deepgram maintains its position as the accuracy and latency leader in voice AI.

Requirements

4–7 years of experience in QA engineering, ML evaluation, or a related technical role with a focus on predictive and generative model and data quality.
Hands-on experience building automated test/evaluation pipelines for ML models and connecting software features.
Strong programming skills in Python; experience with ML evaluation libraries, data processing frameworks (Pandas, NumPy), and scripting for pipeline automation.
Familiarity with speech/audio ML concepts: WER, SER, MOS, acoustic models, language models, or similar evaluation metrics.
Experience with CI/CD integration for ML workflows (e.g., GitHub Actions, Jenkins, Argo, MLflow, or equivalent).
Ability to design and maintain reproducible benchmark environments across multiple model versions and configurations.
Strong communication skills—you can translate model quality metrics into actionable insights for engineering, research, and product stakeholders.
Detail-oriented and systematic, with a bias toward automation over manual process.

Nice To Haves

Experience with model evaluation platforms (Coval, Braintrust, Weights & Biases, or custom evaluation harnesses).
Background in speech recognition, NLP, or audio processing domains.
Experience with distributed evaluation at scale—running evals across GPU clusters or large dataset partitions.
Familiarity with human-in-the-loop evaluation design and annotation pipeline tooling.
Experience with multi-language model evaluation and localization quality assurance.
Prior work in a company where ML model quality directly impacted revenue or customer SLAs.

Responsibilities

Model Evaluation Automation: Design, build, and maintain automated model evaluation pipelines that run against every candidate model before release. Implement objective and subjective quality metrics (WER, SER, MOS, latency/throughput) across STT, TTS, and STS product lines.
Release Gate Integration: Embed model quality checkpoints into CI/CD and release pipelines. Define pass/fail criteria, build dashboards for model comparison, and own the go/no-go signal for model promotions to production.
Agent & Model Evaluation Frameworks: Stand up and operate evaluation tooling (Coval, Braintrust, Blue Jay, custom harnesses) for end-to-end voice agent testing—covering accuracy, latency, turn-taking, and conversational quality and custom metrics across real-world scenarios.
Active Learning & Data Ingestion Testing: Partner with the Active Learning team to validate data ingestion infrastructure, annotation pipelines, and retraining automation. Ensure data quality standards are met at every stage of the flywheel.
Industry Benchmark Automation: Automate execution and reporting of industry-standard benchmarks (e.g., LibriSpeech, CommonVoice, internal production-traffic evals). Maintain reproducible benchmark environments and publish results for internal consumption.
Language & Domain Validation: Build and maintain test suites for multi-language and domain-specific model validation. Design coverage matrices that ensure new languages and acoustic domains are systematically evaluated before GA.
Retraining Automation Support: Validate the end-to-end retraining pipeline across all data sources—from data selection and preprocessing through training, evaluation, and promotion—ensuring automation reliability and correctness.
Manual Test Feedback Loop: Design and operate human-in-the-loop evaluation workflows for subjective quality assessment. Build the tooling and processes that translate human feedback into actionable quality signals for the ML team.