Applied Algorithms Engineer - Information Retrieval

🌟 About You

You like problems with a clear objective
, messy real-world constraints, and lots of room for cleverness.

If you’ve done competitive programming / optimization competitions
, you’ll feel at home here: legal search is basically an optimization game where you trade off quality (F2/NDCG), latency (p95), and cost under strict correctness constraints (citations, traceability, jurisdiction). You’ll build scoring functions, retrieval pipelines, rerankers, and evaluation harnesses; and you’ll ship improvements that users notice immediately.

You enjoy:

• Turning vague user intent into formal signals + algorithms
• Designing fast, low-latency systems under tight budgets
• Running ablations, debugging failure cases, and iterating quickly
• Owning the full loop:
  idea → benchmark → ship → measure

Omnilex is a young, dynamic AI legal tech startup with roots at ETH Zurich
. Our interdisciplinary team (14+ people) empowers legal professionals by building AI systems for legal research and answering complex legal questions; across external sources, customer‑internal documents
, and our own AI‑first legal commentaries
.

As an Applied Algorithms Engineer – Information Retrieval you’ll build the retrieval + ranking + reasoning backbone of our legal research experience.

• Retrieval & ranking beyond the defaults
• Hybrid retrieval (sparse + dense), custom reranking, multi‑stage pipelines
• Domain‑specific workflows (e.g., knowledge graphs, citation‑aware expansions, jurisdiction filters)
• Scoring & features (where algorithms meet relevance)
• Build ranking signals from: citations, authority, recency, jurisdiction, document structure, paragraph/section anchors
• Combine signals into robust scoring functions and reranking strategies
• Query understanding & intent routing
• Classify query intent, detect constraints (“Swiss law”, “latest”, “doctrine vs. case law”), rewrite/expand queries
• Route to the right retrieval strategy with minimal overhead
• Evaluation that actually guides shipping
• Build offline eval sets, define metrics, run quick ablations
• Use production feedback + dashboards to close the loop (what improved? what broke?)
• Search infrastructure & performance engineering
• Tune indices/analyzers/embeddings, manage recall vs. precision, deduplicate near‑duplicates
• Engineer for p95 latency: caching, batching, early‑exit strategies, fallbacks
• LLM‑powered product systems
• Design and ship production‑grade LLM workflows (RAG, tool use, citation‑grounded answers)
• Keep outputs traceable, verifiable, and safe for legal professionals
• Collaboration with domain experts
• Work closely with legal experts to translate pain points into ranking logic
• Document decisions and build playbooks others can extend

• Strong hands‑on experience improving search / retrieval systems in production (hybrid retrieval, reranking, query understanding).
• Proven experience building and deploying LLM‑based products from prototype to production.
• Strong algorithms background (data structures, complexity, graphs, probability/statistics) and practical SQL.
• Proficiency in Type Script/Node.js (our core stack).
• Experience with one or more of:
  Azure AI Search, pgvector/Postgre
  
  SQL, Open Search/Elasticsearch, or similar.
• Familiarity with embedding models + cross‑encoders, and the ability to reason about latency/throughput/quality trade‑offs.
• Ownership mindset, clear communication, bias for action.
• Proficiency in English.
• Full‑time availability.
  Zurich‑based with on‑site presence at least 2 days/week (hybrid).

• Swiss work permit or EU/EFTA citizenship.
• Working proficiency in German
  .
• Experience with evaluation pipelines (human labeling, inter‑annotator agreement, error analysis, AI‑as‑judge—used pragmatically).
• Knowledge of sparse/dense IR methods (BM25 variants, SPLADE, e5/BGE, ColBERT‑style) and semantic reranking.
• Experience operating services (Docker; basic Kubernetes/serverless is a plus).
• Familiarity with Azure / NestJS / Next.js.
• Exposure to legal systems (especially Switzerland, Germany, USA).

• You’ll constantly do “contest‑style”…