AI Engineer - Synthetic Data Generation at Omnilex

Zurich, Zurich, Switzerland -

Full Time

Start Date

Immediate

Expiry Date

06 Jun, 26

Salary

13000.0

Posted On

08 Mar, 26

Experience

2 year(s) or above

Remote Job

Yes

Telecommute

Yes

Sponsor Visa

Skills

TypeScript, Node.js, LLM Integration, Structured Outputs, Data Pipeline Mindset, SQL, PostgreSQL, Drizzle, Testing, Jest, CI/CD, Legal Data Processing, Citation Validation, Cost Optimization, CLI Tooling, Prompt Engineering

Industry

Software Development

Description

🌟 About You Do you get joy from turning messy legal texts into clean, structured, high-quality datasets that actually improve model behavior? Do you like building pipelines where every step is measurable: extraction quality, citation correctness, dedup rate, cost per item, throughput, and regression stability? Are you comfortable shipping pragmatic tooling (CLIs, validators, tests) around LLMs without hand-waving away edge cases? If so, we’d love to hear from you. 🚀 About Omnilex Omnilex is a young dynamic AI legal tech startup with its roots at ETH Zurich. Our passionate interdisciplinary team of 10+ people is dedicated to empowering legal professionals in law firms and legal teams by leveraging the power of AI for legal research and answering complex legal questions. We already stand out with handling unique challenges, including our combination of external data, customer-internal data and our own innovative AI-first legal commentaries. 🧬 Your Mission: Synthetic Data for Legal AI As an AI Engineer – Synthetic Data Generation, you will build and own pipelines that generate retrieval-ready and evaluation-grade synthetic datasets from real legal sources (court decisions, statutes, commentaries) across languages and jurisdictions, while keeping quality high and costs controlled. Tasks 🛠️ Your Responsibilities Build multi-step generation pipelines (10+ steps): From DB selection → pseudonymization → extraction → translation → normalization → deduplication→ validation → classification → rating → export. LLM integration, production-grade: Design robust prompt suites for extraction, translation, classification, and rating; enforce structured JSON outputs; handle retries, partial failures, and weird model behavior. Quality assurance & filtering: Implement scoring systems (multi-criteria, consistent rubrics), dedup/near-dup suppression, and deterministic validators (especially for citations). Citation processing at legal-grade precision: Extract, normalize, and validate citations across languages and formats (e.g., Art. 336c Abs. 1 OR, BGE 137 III 266 E. 3.2), including abbreviation mapping and normalization rules. Cost & throughput optimization: Use batch APIs where appropriate, tune reasoning effort, control concurrency, count tokens, and keep runs cost-efficient (without sacrificing quality). Developer tooling & CLI workflows: Build CLIs with progress tracking, configurable concurrency, and solid ergonomics for long-running jobs. Testing across levels: Write unit/smoke/integration tests for pipelines and validators (including mocked LLMs where sensible and real API runs where needed). Cross-team collaboration: Work closely with legal experts to define what “good” looks like for exam questions/commentaries, and translate that into measurable QA checks. Requirements ✅ Minimum qualifications Experience building backend/data tooling with TypeScript/Node.js (strict typing, generics, async patterns). Hands-on experience integrating LLM APIs (OpenAI/Anthropic or similar), including structured outputs (JSON), prompt iteration, and failure handling. Strong data pipeline mindset: ETL workflows, transformation steps, validation, and reproducibility. Solid SQL/PostgreSQL skills and experience with an ORM (bonus if Drizzle). Experience writing reliable tests (e.g., Jest) and maintaining CI-friendly pipelines. Fluent English; willing to work hybrid in Zurich (on-site at least two days/week), full-time. 🎯 Preferred qualifications Familiarity with the Swiss legal system (court structure, citation norms, multilingual legal terminology). Working proficiency in German; plus French/Italian is a strong advantage. Experience with batch processing and cost-aware LLM operations (token budgeting, batching strategy, caching, early-exit). Practical text processing skills: regex-heavy parsing, dedup/near-dup detection, similarity search (e.g., BM25 / MiniSearch). Familiarity with our environment: Yarn workspaces/monorepos, NestJS, and pragmatic CLI tooling. Benefits 🤝 Benefits Direct impact: Your datasets will directly shape model quality and evaluation reliability in legal research and reasoning. Autonomy & ownership: Own the synthetic data pipeline end-to-end; prompts, validators, QA, exports, and cost controls. Team: Work with a sharp interdisciplinary group at the intersection of AI, engineering, and law. Compensation: CHF 7’000–11’000 per month + ESOP, depending on experience and skills. We’re excited to hear from candidates who love building robust, cost-aware LLM pipelines and care about precision (especially when citations and multilingual legal nuance matter). Apply today by pressing the Apply button.

Responsibilities

The engineer will build and own multi-step pipelines (10+ steps) to generate retrieval-ready and evaluation-grade synthetic datasets from real legal sources across languages, focusing on high quality and controlled costs. Responsibilities include designing robust LLM prompt suites for various tasks, implementing rigorous quality assurance, and ensuring legal-grade precision in citation processing.