About the RoleLanguage Translation is one of Sanas's most exciting and fastest-growing product lines. We're looking for a Research Engineer who can both set technical direction and get deep in the modeling work - someone who owns translation quality end-to-end across language pairs and drives the fundamental research challenges unique to real-time simultaneous interpretation.
Job DescriptionTranslation quality & modeling- Own and drive improvements to translation accuracy across Sanas's supported language pairs, with a focus on conversational, spoken-language domains.
- Design, train, and evaluate neural MT models - from fine-tuning large multilingual models to building targeted components for low-resource or high-priority language pairs.
- Develop and maintain rigorous evaluation pipelines using both automated metrics (BLEU, COMET, chrF) and human evaluation frameworks calibrated to real-world enterprise use cases.
- Identify the highest-leverage research bets - data augmentation, domain adaptation, quality estimation, terminology consistency - and execute on them with measurable quality gains.
Simultaneous interpretation & delimiter modeling- Lead research and development of Sanas's delimiter model - the component that determines optimal segmentation points in streaming speech for real-time translation output.
- Develop methods to handle speech disfluencies, sentence fragments, and incomplete utterances gracefully in a streaming translation pipeline.
- Collaborate closely with the speech and inference engineering teams to ensure translation components meet strict real-time latency budgets in production.
Research direction & technical leadership- Define and maintain a research roadmap for MT and simultaneous interpretation, prioritizing work that moves production quality metrics.
- Stay at the frontier of MT research - track and evaluate relevant work - and translate (haha) relevant advances into practical improvements at Sanas.
- Mentor and technically guide other engineers working on translation-adjacent problems across the ML org.
Data & infrastructure- Identify, source, and curate training data for MT and delimiter modeling - including parallel corpora, synthetic data generation, and speech-aware augmentation strategies.
- Instrument model quality monitoring in production to detect degradation across language pairs and trigger targeted retraining cycles.
Qualifications- 3+ years of experience in machine translation, NLP, or multilingual modeling research - with a track record of measurable quality improvements in production systems.
- Deep familiarity with neural MT architectures: sequence-to-sequence models, Transformer variants, and large multilingual models.
- Hands-on experience with simultaneous or streaming translation, including segmentation and low-latency decoding strategies.
- Strong command of MT evaluation methodology - automated metrics, human evaluation design, and error analysis.
- Proficiency in Python and deep learning frameworks (PyTorch preferred)
- Demonstrated ability to set a research agenda, execute independently, and communicate findings clearly to technical and non-technical stakeholders.
- Fluency in English plus working proficiency in at least one non-English language is a strong plus.
Bonus- Experience with speech translation (end-to-end or cascaded) and speech-aware MT pipelines.
- Familiarity with on-device or edge-optimized model deployment for low-latency inference.
- Prior work on low-resource language pairs, domain adaptation, or terminology-constrained translation.
- Published research at ACL, EMNLP, NAACL, INTERSPEECH, or equivalent venues.
