Principal AI Researcher 

At Phrase, we help open the door to global business by providing the world's leading Language Intelligence Platform. 

The Phrase Platform combines AI, agentic orchestration, and a headless, API-first architecture in one composable system. Beyond translation, it orchestrates and adapts content to culture, audience, channel, brand voice, and intended outcome in any language, for every audience. It applies the context that makes content perform in every market: quality standards, glossaries, prior translations, and cultural nuance. Every team, in every region, can ship content that is on-brand, on-point, and ready for any audience.

Phrase gives enterprises the intelligence to automate workflows, the freedom to connect their own tools and engines, and the control to govern global content at scale.

The AI Research team sits at the core of Phrase’s product differentiation. We build, train, and evaluate proprietary translation models; we design the agentic workflows that orchestrate translation quality end-to-end; and we run the research cadence that feeds product development with grounded, evidence-based decisions.

The team’s current research programmes span: fine-tuning LLMs for domain-specific machine translation, multi-agent evaluation pipelines, automated quality profiling from style guides, and active learning through feedback loops on real customer data. These are not exploratory proofs of concept; they are live systems serving production traffic.

The next phase requires someone who can take ownership of the most technically demanding of these streams, extend them, and help chart the research direction for the team as a whole.

What you’ll be responsible for:

Core Research

• Lead design, training, and evaluation of LLMs for translation and language quality tasks, including work with fine-tuning techniques such as LoRA and DPO, and instruction-tuned models at various scales.

• Design and implement robust evaluation frameworks for translation quality, moving beyond automatic metrics (BLEU, TER, ChrF, MQM, COMET), LLM-as-judge approaches, and hybrid evaluation pipelines.

• Architect and evaluate complex agentic workflows for NLP problems: multi-step reasoning, tool use, structured output generation, and orchestration across multiple model providers.

• Take technical ownership of the team’s flagship model development programme, from data curation and training pipeline design through to production integration and ongoing evaluation.

• Design and run experiments using ML pipelines, maintaining reproducibility and clear documentation of results and decisions.