AI Engineer

Frontier's platform turns execution data into approved action — detecting drift, exposing risk, and orchestrating the next step for complex operations teams. The AI Engineering team in Sofia owns the capabilities that make this work: the agents, models, and pipelines that sit behind our product.

The Role

You are an AI engineer based in Sofia, owning AI capabilities end-to-end — from problem framing through prototype, productisation, and post-launch monitoring. Your work spans LLM-based agent systems and model development.

You work alongside the Sofia research team, adopting the modeling patterns they establish and contributing back the production-readiness rigor that turns research artefacts into shipped capability. You set the example through the quality of your work — the standards of evaluation rigor, safety, and operational discipline you bring are the ones the rest of the team follows — and you serve as a senior engineering anchor in the EU time zone, providing review, guidance, and unblocking across our distributed team.

Tech Stack

• Backend API: Python, FastAPI, Pydantic, SQLModel
• AI agents: LangGraph, LangChain, Pydantic AI
• ML: PyTorch, scikit-learn, XGBoost, NeuralProphet
• Data processing: Polars, Pandas, Databricks
• PDF processing: PyPDF, PyPDFium2, PyMuPDF
• Browser automation: Playwright
• Database & cache: PostgreSQL, MariaDB, Valkey
• Observability: Logfire, Sentry, Datadog

What We're Looking For

• 5+ years in applied AI / ML in production — Systems with real users behind them, not only research papers or personal projects.
• Agentic products — Strong experience building LLM applications and multi-agent workflows in production, with deep familiarity with prompt design, tool use, RAG, memory architectures, evaluation, and the failure modes that only appear at scale. Hands-on with LangGraph, LangChain, or comparable orchestration frameworks.
• Model development — Practical experience training, fine-tuning, and evaluating models; LoRA and PEFT techniques; dataset curation; and the judgement to know when fine-tuning is the right tool versus prompting or retrieval.
• Evaluation — You build the eval harness alongside the feature: dataset curation, metrics, LLM-as-judge, agent trajectory analysis, and production quality monitoring.
• Data fluency — Comfortable with SQL against large relational databases and with the realities of extracting signal from messy production data.
• Production mindset — Familiarity with prompt injection defenses, PII handling, output validation, jailbreak prevention, rate limiting, runaway-cost guards, and circuit breakers for agent loops in a multi-tenant SaaS environment.
• Computer vision — Experience with image segmentation, object detection, or polygon extraction is a strong advantage.
• Communication — Clear, precise PR descriptions, technical proposals, evaluation reports, and architecture notes.