Pinnipedia is a new Berlin startup building a cloud platform that automates and assists the creation of **audit-ready IT-security concepts** (e.g., BSI-Grundschutz, C5). We’re IGP-funded (2025/26) and co-develop with FU Berlin and pilot users from industry and security consulting.
We’re looking for a freelance **AI Engineer** to turn messy inputs into structured knowledge and reliable answers.
**Your Mission** -Own the end-to-end pipeline that turns unstructured documents into a validated, queryable knowledge graph. Accountable for extraction quality, graph integrity, and the data layer that backs the product's read path.
## Tasks
• **LLM extraction pipelines** -document chunking, property and relationship extraction, cross-chunk reconciliation, gap detection. Built with structured-output LLM agents orchestrated by durable workflows.
• **Knowledge graph** -schema design as typed Pydantic models, Cypher access patterns and indexing strategy, graph operations, schema evolution and migration. Scope ends at the graph boundary: API contracts and query abstractions exposed to consumers belong to the full-stack engineer.
• **Deterministic rule engines** -table-driven evaluators for cases where code beats LLM judgment; clear contracts between deterministic and probabilistic components.
• **Data validation & quality** -schema enforcement, required-property contracts, audit trails, eval harnesses (expert review, unsupervised checks, synthetic fixtures, LLM-as-judge).
• **Live data ops** -backfills, coordinated migrations across relational + graph stores, observability on extraction throughput and quality, incident response.
## Requirements
**Must-have**
* 5+ years shipping data/AI systems to production with real customers -has been on-call for live pipelines and knows what breaks at 2am.
* Strong Python (typed, modern) and SQL. Comfortable with PostgreSQL under load.
* Production experience with at least one graph database (Neo4j preferred; Neptune, ArangoDB, TigerGraph acceptable) -schema design, query tuning, not toy use.
* Production LLM pipeline experience: structured output, agent orchestration, prompt and version management, evaluation frameworks. PydanticAI, LangChain, DSPy, or Instructor all welcome.
* Durable workflow orchestration in production (DBOS, Temporal, Airflow, Prefect, Dagster).
* Test-first discipline -integration tests against real datastores (Testcontainers or equivalent), not mock-heavy unit tests.
* Fluent English skills.
* Freelance status.
**Nice-to-have**
* Experience with regulated, compliance-driven, or standards-heavy extraction domains (legal, medical, financial, security/audit).
* Designed deterministic evaluators alongside LLM components and knows when to reach for which.
* Contributions to data contracts, schema governance, or ontology work.
* German language skills.
## Benefits
**Remote, full-time** with flexible scheduling. On a freelance basis. **CET (Berlin) timezone availability expected.**
Possibility of relocation if successfull work relationship is achieved after a period of time.
**Competitive salary: 75.000–85.000 €** base (premium for exceptional senior profiles).
Small, focused team; direct collaboration with the Product Owner and Full-Stack Engineer.
Modern tooling, real ownership, and a learning budget for role-relevant training.
Impact: help SMEs meet rising security requirements with less friction.
**Apply on JOIN** with your CV (PDF) and a short note (max **200 words**) describing **how you would design a KG-backed RAG pipeline** (ontology scope, indexing, retrieval, and evaluation you’d use).
**Process:** 20-min intro → 90-min practical (graph modeling + retrieval evaluation) → 45-min team chat → references. We review applications within **5 business days**.