27.10.2026 aktualisiert
Premiumkunde
100 % verfügbarPython Backend & AI Application Engineer | LLM | RAG | Agentic Workflows
Buttenwiesen, Deutschland
Nur Remote
BA Communication & Cultural Sciences, Universidade Catolica PortuguesaÜber mich
Software engineer turned AI engineer: I build and deploy LLM systems end-to-end, from RAG to agentic workflows. Stack: Python, FastAPI, LangChain, LangGraph, MCP/OAuth 2.0, Pinecone, Docker, Google Cloud Run, GitHub Actions. Former tech lead. German (B2), English (C1).
Skills
AI AgentsRAGLangChainPythonWorkflowsLarge Language ModelsPrompt EngineeringBackendFastAPIDocker
LLM & RAG Pipeline Development
Design and deployment of LLM-based systems end-to-end, including RAG pipelines, prompt engineering, agentic workflows, and NLP at scale using LangChain, LangGraph, OpenAI API, Groq API, and sentence-transformers.
Python Backend Engineering
Development of robust Python backends with FastAPI, asyncio, REST APIs and SQLite, including SSE streaming endpoints and async fire-and-poll patterns for long-running jobs, with a strong focus on end-to-end ownership from prototype to production. Occasional frontend work using React/TypeScript when a project calls for it.
AI/ML Fine-Tuning & Evaluation
Fine-tuning of large language models using LoRA, QLoRA and Hugging Face PEFT. Design and maintenance of LLM-as-judge evaluation sets with custom rubrics, tracked in LangSmith across iteration checkpoints, including adversarial and cross-lingual cases; BERTScore and ROUGE for retrieval and summarisation quality.
EU AI Act Transparency (Art. 50)
Implementation in a live LLM system: AI-disclosure surfaces, a dedicated identity intent in the agent router resolved without a model call and answering in the user's own language, adversarial test cases for identity-disclosure and prompt-exposure, machine-readable generation provenance as the basis for Art. 50(2), and a documented feature-tripwire register that forces re-assessment before shipping.
NLP Pipelines
Design and implementation of full NLP pipelines including text preprocessing, embedding generation, semantic clustering, and sentiment classification at scale using Hugging Face Transformers, scikit-learn, and Pandas.
Vector Search & Embeddings
Retrieval architecture using Cohere Embeddings and Rerank, Pinecone, and sentence-transformers for cross-lingual similarity search and re-ranking of retrieved candidates.
Deployment, Containerisation & CI/CD
Containerised deployment with Docker and Docker Compose on Google Cloud Run, including non-root images and nginx reverse proxy; batch workloads as Cloud Run Jobs. Automated pipelines and required-check CI gates via GitHub Actions covering unit tests (pytest) and Playwright end-to-end runs. FastAPI, Streamlit, MLflow, Git.
Tech Lead & Agile Practices
Tech lead responsibilities including architectural decisions, code reviews, unblocking engineers, maintaining quality standards, and translating business requirements into technical specifications in agile environments.
Erfahrung
Berufserfahrung0 - 5 Jahre
Freelancer seit2026
Sprachen
DeutschgutEnglischverhandlungssicherPortugiesischMutterspracheSpanischGrundkenntnisse
Projekthistorie
Built a Model Context Protocol server with OAuth 2.0 as a first-class concern, exposing a mock purchase-to-pay process to AI agents.
- Implemented the full authorization code flow against Keycloak: client identity via a publicly hosted metadata document instead of dynamic client registration, PKCE, no client secret anywhere in the project, token audience validated at the resource server, and signatures verified locally against the published key set.
- Every call is decided per request by granted scope intersected with roles the server resolves itself, so a token is a ceiling on what an application may do on a user's behalf rather than a statement of what that user is allowed to do — a tool the caller lacks the role for is absent from the listing rather than refused on call.
- Modern protocol revision 2026-07-28 over Streamable HTTP, run stateless as two replicas behind nginx so statelessness is falsifiable rather than asserted.
- Authorization policy held as data: a 33-row decision matrix driving the tests, the seeded fixtures and the published documentation from one source, plus a 34-row attack suite with 19 rows citing the specification clause they defend, each in enforced bijection with the test that falsifies it.
- 533 automated tests and seven CI checks, one of which completes a real authorization code flow against a running stack on every pull request. Proven against three independent clients including Anthropic's own shipped client.
Designed a RAG-style retrieval pipeline using sentence-transformer embeddings and structured LLM evaluation against a target profile. Built an async ingestion layer with pluggable source adapters, developed structured LLM evaluation framework with scoring calibration, and engineered an idempotent SQLite-backed pipeline with deduplication and feedback-driven re-ranking. Developed ~300 pytest tests ensuring end-to-end system reliability; deployed on GitHub Actions as a daily automated pipeline.
Originally built during an Ironhack bootcamp project, then continued and substantially extended independently into a production system: multilingual RAG over a corpus of 50 science video transcripts (EN/ES/DE/FR/PT), with a FastAPI backend and a React/TypeScript SPA deployed as separate Google Cloud Run services behind nginx, SSE token streaming and timestamped source pills.
Retrieval on Cohere embed-multilingual-v3.0 with asymmetric input types, plus a Cohere Rerank v3.5 cross-encoder re-scoring 10 over-retrieved candidates down to 3. Agentic orchestration with LangGraph. Corpus pipeline as a Cloud Run Job backed by GCS with idempotent, resumable steps; all three services containerised as non-root images; CI gates every pull request on four required checks incl. Playwright end-to-end.
Implemented the EU AI Act Article 50 transparency obligations ahead of the 2 August 2026 application date: AI-disclosure surfaces, a dedicated identity intent in the agent router resolved without an LLM call and answering in the user's own language, unit tests plus five adversarial eval cases, machine-readable generation provenance for Art. 50(2), and a documented feature-tripwire register.
Iteratively improved across 7+ evaluation checkpoints against a 46-case LLM-as-judge eval set (10 adversarial cases reviewed manually) tracked in LangSmith — best checkpoint mean 4.67/5 over the 33 scored cases.