AI Researcher

Frontier builds Execution Intelligence — AI that helps complex operations teams detect drift, map exposure, and keep high-stakes work moving. Our Sofia AI research team sets the modeling patterns and practices the rest of the organization builds on.

The Role

You will join our AI research team in Sofia — the group that sets the modeling patterns and research practices the rest of the AI organization builds on. You take ambiguous problems in language understanding, retrieval, knowledge representation, and forecasting, and turn them into methods that work: reading the literature, designing experiments, building prototypes, and proving out approaches before they become product.

You design the evaluation frameworks that tell us whether a model or agent is actually good, and you partner with AI engineers to hand off research artefacts in a state they can productionise. You raise the research bar through the rigor of your experimental design and the clarity with which you communicate findings to a distributed team.

Tech Stack

• Core: Python, PyTorch, Hugging Face, scikit-learn
• Agent prototyping: LangGraph, LangChain, Pydantic AI
• Data & experimentation: Polars, Pandas, Databricks
• Database: PostgreSQL, MariaDB
• Observability: Logfire, Sentry, Datadog

What We're Looking For

• Research background — MSc or PhD in CS, ML, or a related field, or equivalent demonstrated research experience. Publications, competition results, or shipped research are a plus.
• NLP / LLMs — Deep familiarity with transformer models, fine-tuning, retrieval-augmented generation, prompting, and the current research landscape.
• Knowledge representation — Experience with knowledge graphs, structured retrieval, or hybrid neural-symbolic approaches is a strong advantage.
• Evaluation — You design datasets, define metrics, and build LLM-as-judge and trajectory-analysis approaches. You treat evaluation as a first-class research output, not an afterthought.
• From paper to prototype — You can read a recent paper and reproduce or adapt it in code, and you know when a simpler baseline is the honest answer.
• Data fluency — Comfortable with SQL and with extracting training signal from messy, real-world data.
• Communication — Clear written research: proposals, experiment write-ups, and findings that a distributed team relies on.