Neo4j Experts in Germany

• Fully on-premises LLM document-examination platform for a compliance-critical banking domain: agentic LangGraph pipeline with deterministic verification, every AI judgment structured and source-anchored; ~960 automated tests, zero data egress
• GPU throughput engineering (quantized serving, speculative decoding, prefix caching): 9.5x extraction speed-up, 500+ multi-document case files per day on a single A100
• AI-native EDI/EDIFACT integration platform (~116k LOC Java 25 / Spring Boot 4, 1,900+ tests): LLM-drafted partner mappings machine-verified before go-live (DFDL conformance, field-coverage checks, dry runs), ~99.5% byte match on real customer files — replacing weeks of manual mapping per partner

Multi-Agent Document-Generation Platform

Problem:

Turn a single prompt into a well-researched, structured and cited document, while keeping the user in control: every stage can be reviewed, corrected or revisited before the system moves on.

Architecture:

A multi-agent system orchestrated by a central controller. More than 20 specialised pydantic-ai subagents (understanding, structuring, planning, research, writing, assembly, cross-reference check, glossary, citation, reference table and more) pass results forward through a shared plan persisted in Postgres (asyncpg); a runner layer dispatches each task by name to its agent, and mechanical steps such as formatting, table of contents and list of figures are deterministic code instead of LLM calls. Provider (Anthropic or OpenAI) is a configuration setting, and responses stream over Server-Sent Events.

Hosting & Infrastructure:

Stateless container run as a Scaleway Serverless Container on a private network with managed PostgreSQL, object storage for finished documents and assets, and centralized log shipping (Loki / Scaleway Cockpit), provisioned with OpenTofu modules and deployed through GitLab CI; usage is metered per API key.

My Contribution & Key Decisions:

• Central decision controller: a satisfaction classifier turns the user's free-text feedback into a verdict (keep working, advance, cancel or revise an earlier task), and a middleware applies it to the plan. Plan progression supports jumping back to earlier tasks and re-running the stages after them.
• Research and writing fan-out: planning produces a plan from which a research-design agent generates individual research tasks and a writing-design agent generates individual writing tasks, each linked to the research it needs. Every task runs in its own dedicated agent with its own conversation history.
• Self-review and human-in-the-loop: after each step a critical-thinking agent reviews the output; the user can accept, refine, or ask questions about a critique before the plan advances.
• MCP integrations: two remote MCP servers, web search and internal document search, connected through pydantic-ai with retries and long-running tool-call timeouts. Classifier agents decide per task whether web search is needed, and a small model condenses raw search results into structured findings before they reach the writing agents.

Usage Metering:

Clients authenticate with an API key validated against a key-management service; consumption is charged in credits calculated from token usage per model plus a markup, and exhausted quota is answered with HTTP 429.

• This project involved extraction of intelligence data to support report writing for a company that provides geopolitical, global, commercial intelligence. The data have been gathered from a number of resources (interview transcripts, online data, internal documents), and then a knowledge base has been build from it. This was the basis of a complex RAG system, that was evaluated against a golden dataset. Agents have been used to find out the contradicting intelligence, the statements supporting each other, and to store back the generated knowledge.

Used: Python, RAG, LangGraph, LangChain, deepeval, MCP

• Building up the Enterprise Architecture department. Preparing the convergence of IT / OT.
• Mapping the global business application landscape (6 countries).
• Concept for a graph model of the application landscape. Storing the results in the open-source graph database Neo4J. Various reports in the database query language CYPHER and the visualization component BLOOM. Automatic generation of IT architecture diagrams.
• Development of an application taxonomy to identify redundant applications.
• Applying Gartner's Software Portfolio Management TIME model (Tolerate, Invest, Migrate, and Eliminate) for portfolio management.
• Enterprise Architecture Governance: establishing architecture guidelines for documenting measures and decisions.
• Process modeling according to Porter (Value Chain).
• Various architecture approvals for change requests and documentation in Architecture Decision Records (ADR).
• Working on the RACI matrix for various processes.
• Working on business capabilities as preparation for a LeanIX rollout.
• Contributing to the concept for IT/OT convergence.
• Establishing architecture guidelines in a very pragmatic corporate culture. Knowledge about applications and responsibilities is spread across the world.

Independent, part-time open-source engineering focused on build-time tooling for Next.js, React, MDX, and JavaScript/TypeScript compiler pipelines.

• Built next-slug-splitter to optimize content-driven Next.js applications. It analyzes MDX content at build time, resolves component usage, and generates route-specific handlers so pages avoid sharing the full catch-all component bundle.

• Created supporting plugins and utilities for scoped MDX transformations, nested component dependency resolution, compile-time refinement, safe ESTree evaluation, and object-graph diffing.

• Own architecture, API design, implementation, automated testing, npm publishing, documentation, demos, and performance benchmarking.

Building blocks:

• remark-scoped-mdx: Context-aware AST transformations with nested scope isolation, typed component registries, and prop inference.

• recma-component-resolver: Dependency-graph analysis and selective component forwarding across nested MDX includes.

• recma-static-refiner: Build-time prop extraction, schema validation, derivation, and pruning.

• estree-util-to-static-value and object-graph-delta: Safe static evaluation and deterministic, cycle-safe structural diffing.

Tech Stack:

• Frameworks: TypeScript · Next.js · React · MDX

• Compiler tooling: Unified · Remark · Recma · MDAST · ESTree · ts-morph · esbuild

• Competencies: Static analysis · AST traversal and transformation · dependency graphs · code generation · schema validation · route and bundle splitting

• Tooling: Vitest · tsup · npm · performance benchmarking

• Portfolio development & customer acquisition
• Portfolio development (RAG, NLP fine-tuning, process automation with n8n) and active customer acquisition
• Positioning: GDPR-compliant, locally hosted AI solutions for SMEs

• Developing a model-based AI where natural language is interpreted as formal relations.
• Natural language terms are not considered rigid but fluid and can be negotiated in a context so meaning resolves by iteratively specifying.
• Develops all kinds of model solutions.
• Backed by natural language and data annotation.
• Requirements to code and other solutions.

Statistical Forecasting Algorithm

• Improvement of an statistical probabilistic forecasting algorithm for sales + evaluation
• Migration from R/On-premise to Python/Snowflake
• Productionalization on Snowflake in cooperation with data engineers & DevOps

Monitoring Dashboard

• Data engineering for preparation & provisioning of necessary data/resources on Snowflake
• Development & deployment of a Streamlit dashboard in Snowflake

ML-based Probabilistic Forecasting on Vertex AI

• Development of a ML-based probabilistic forecasting algorithm from scratch
• Implementation of MLOps pipeline in Kubeflow on Google Cloud Vertex AI

Tech Stack: Python, Snowflake/Snowpark, R, Streamlit, Gitlab/Gitlab CICD, Terraform, Google Cloud, Vertex AI (aiplatform SDK, gcloud CLI, feature store, model registry, etc), kubeflow

Designed and developed an Azure-based AI and data platform transforming distributed project documents into searchable, structured knowledge for an enterprise AI assistant. Combined incremental document processing, hybrid keyword/vector retrieval, knowledge graph extraction, and conversational AI for Microsoft Teams and Microsoft 365 Copilot. Implemented Python document-processing pipelines for PDF, Word, Outlook and text files; Azure Cosmos DB lifecycle management; Azure AI Search; Azure AI Foundry and Azure OpenAI; identity-aware retrieval with Microsoft Entra ID; project authorization; a TypeScript Microsoft Agents SDK agent; Pulumi infrastructure; resilient scheduling, retry and diagnostic tooling. Technology stack: Python, TypeScript, FastAPI, Pydantic, Azure AI Foundry, Azure OpenAI, Azure AI Search, Azure Cosmos DB, Blob/Table Storage, Container Apps/Jobs, Service Bus, Key Vault, Entra ID, Docker, Pulumi, pytest, Microsoft Teams and Microsoft 365 Copilot.

"Reasoning Models: From LLM Foundations to NOC Automation."

UEF 3.0 · Semantic Government Overlay (SGO) · Autonomous Systems (UAS / dual use)

• Designed: Semantic Government Overlay (SGO) – AI-guided administration without replacing existing specialist procedures. Read-only semantic layer over registers and specialist processes based on the Federal Information Management (FIM). Decision authority remains with the case worker (architecture principle).
• Developed: Reference architecture with source-backed, derived statements (Executable Ontologies OWL/RDF/SHACL). Technically guaranteed purpose limitation and no-write-path principle in specialist data – auditable, without a central data pool.
• Anchored: Regulation as a design principle: EU AI Act (high-risk obligations for public-sector AI, fundamental rights impact assessment under Art. 27), GDPR, NIS2, and administrative automation limits (§ 35a VwVfG, § 31a SGB X) as technical control points in the architecture.
• Created: Methodical tool for pilot organizations: data pipeline assessment (phase 0), compliance blueprint, and management summary as a decision-ready package for public administration.
• Specified: UEF 3.0 as a successor architecture to TOGAF – decision paper, canonical ontology, six-layer architecture, read/actuate boundary, federation registry, terminology concordance, and release delta as a closed specification status.
• Architected: AI-native mission OS for autonomous UAS and ground robotics as a tactical layer on top of a separately approved autopilot. Run-time assurance according to ASTM F3269-21 (Simplex pattern): the verified safety controller keeps authority, the AI function provides suggestions.
• Designed: Three-tier architecture – Tier 0 autopilot with 650 Hz flight control on RTOS, Tier 1 AI OS with semantic world model and multi-agent cluster, Tier 2 swarm and ground mesh. Zenoh as the primary fabric, MAVLink as the only authenticated command path (single writer). Result: graceful degradation – loss of the mission, not of the aircraft.
• Secured: Two-gate chain on the read/actuate boundary – governance gate (can-question: AI Act risk class per actuation, enforced human oversight under Art. 14, immutable log) before the RTA safety monitor (is-it-correct question: flight envelope, geofence, energy reserve) with revert to the baseline controller.
• Anchored: Dual-use architecture with common core and build-time fork instead of runtime switch. Three separate legal levels: civil variant – UAS under the EASA Basic Regulation (EU) 2018/1139 with the limited applicability under Art. 2(2) of the AI Act, ground robotics under the Machinery Regulation 2023/1230 with the full high-risk obligation chain, Cyber Resilience Act for both; unarmed carrier variant as defense material under AWG/AWV and Dual-Use Regulation 2021/821 (BAFA approval); armed variant under KrWaffKontrG. Each variant lives under exactly one dominant legal regime. Evidence base: AI BOM, SBOM, and complete data lineage.
• Analyzed: System analysis and realignment of grown engineering system landscapes. Approach concept for consolidation without migration – semantic layer over the existing sources instead of data transfer. Result: decision-ready implementation concept including an evaluation model for the target architecture.

• Build and operate an end-to-end MLOps platform on Azure ML and Kubernetes (Kubeflow) for the automated deployment, monitoring, and scaling of forecasting models (including Temporal Fusion Transformer, Informer, Autoformer).
• Implement CI/CD pipelines in Azure DevOps for the full ML lifecycle – from resource provisioning (Terraform), data transformation (Hugging Face Datasets, Pandas, PyTorch, CUDA cluster) through training and evaluation to model registry and endpoint deployment.
• Integrate MLflow for experiment tracking, model versioning, performance monitoring, and automated registration in the Azure Model Registry.
• Develop and containerize PyTorch training jobs (Azure Notebook, Jupyter Notebooks) for price and time series forecasting (PFC models) with automatic rollout via Azure ML Endpoints and REST/gRPC interfaces, Docker containerization, secured with OAuth 2.0.
• Set up monitoring and alerting mechanisms (Prometheus, MLflow Metrics), log centralization, and cost monitoring.
• Automate infrastructure provisioning and model deployment using Terraform, Helm, and Azure CLI; connect to existing market data systems and event pipelines.
• Migrate existing workloads and databases (IONOS → Azure, MongoDB) with integration into central MLOps workflows and internal networks.
• Extend the platform with LLM-based tools (LangChain, LangServe) to integrate GPT-based analysis modules into existing Spring Boot services for market anomaly detection and automated reports.
• Analyze and architect a software solution to process large volumes of data efficiently (>3000 messages/sec.) (market data store).
• Spring Boot / Java 21 container development with RabbitMQ for distributing stock market data via MongoDB (Kubernetes) with fast storage of data in Redis RMaps, deduplication, forwarding messages to Read Model queues, and building Read Models for UI display in MongoDB.
• Integration of RESTHeart to create a REST API for MongoDB.
• Build an Angular frontend to simplify data queries and master data maintenance.
• Agentic coding with remote and local LLMs (Claude Sonnet, Ollama Qwen) and MCP servers.
• Develop Python scripts for transforming and cleaning incoming stock market data (Pandas, scikit-learn).

• Architect end-to-end AI/ML solutions, with a strong focus on advanced LLM applications, autonomous agentic AI workflows using MCP toolset, vLM embedding computations, and robust agentic RAG pipelines to drive real-world business impact for clients across diverse sectors.
• Designed intelligent agentic AI workflows to automate knowledge retrieval and reasoning for B2B clients, accelerating business impact.

• Partner directly with C-level leadership (CEO, CAIO, CTO) on architecture, OKR strategy, and cross-team roadmap prioritization, translating strategic goals into structured engineering requirements.
• Surfaced and mapped technical debt across the entire organization with C-level leadership and co-defined a prioritized remediation strategy, balancing debt paydown against feature delivery.
• Led code reviews and technical standards across the team, fostering a mentor-first environment with two-way feedback dialogue — pairing on complex pipeline work and unblocking junior engineers on async architecture patterns.
• Re-architected the AI companion's core processing pipeline from synchronous to asynchronous with a queue-based worker architecture, enabling horizontal scalability and cutting upload processing time ~4x (from ~22s to 5–10s) while improving response accuracy.
• Designed an AI-driven document intelligence workflow with automatic multi-document classification, per-document summarization, and relevance guardrails for the patient care journey.
• Built a unified patient memory system (short- and long-term context) bridging the document vault and chatbot into a single bidirectional, context-aware platform.