Pydantic Experts in Germany

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.

• Led end-to-end delivery of ML/AI applications for the QuantistryLab simulation platform, from requirements definition and solution design to development, testing, deployment, and production support, coordinating efforts across back-end, DevOps, and ML teams
• Designed and implemented a compound search system with an open-weights LLM, the Pydantic AI agent framework, and the Model Context Protocol (MCP), exposing it through FastAPI endpoints
• Deployed and configured vLLM-based LLM serving on Kubernetes
• Built CI/CD pipelines and Helm charts to automate deployments across staging and production environments
• Instrumented the search system with OpenTelemetry spans and traces, integrating with the Honeycomb observability platform to monitor component-level latency and functional errors
• Developed a custom evaluation dataset and implemented evaluations for retrieval relevance, answer correctness, tool-call accuracy, and end-to-end task success
• Reduced infrared spectrum prediction time by 83% (from 60 to 10 minutes) by replacing a quantum-mechanical model with a pretrained graph neural network; fine-tuned the model using PyTorch, built and containerized the inference pipeline with Docker, and integrated it into the platform
• Implemented an MLflow model registry with MinIO-backed artifact storage
• Advised sales and science teams on ML/AI applications for prospective client projects, providing technical expertise to shape proposed solutions and prototypes

Technologies: AWS, Hetzner, Kubernetes, Helm, Docker, vLLM, Pydantic AI, MCP, Qwen, Llama 3, MLflow, MinIO, GitHub Actions, Python, FastAPI, OpenTelemetry, Honeycomb, PyTorch, Polars, pytest

• Designed and developed a multi-tenant SaaS platform enabling organizations to build their own knowledge bases and chat with brand-customized AI assistants (white-label approach with dynamic branding per organization).
• Implemented a scalable RAG architecture with a GPT-4o tool-use loop, hybrid semantic search, and strict tenant isolation at database and search index level.
• Built persistent, project-like chat sessions including a streaming API (SSE), multilingual support, and speech input/output (STT/TTS).
• Delivered the cloud infrastructure as Infrastructure-as-Code, fully automated per-customer CI/CD pipelines, and an onboarding process for new tenants.

Technologies used: Python, FastAPI, Pydantic (v2 noted), Next.js, React, TypeScript, Tailwind CSS, OpenAI / LLMs (GPT-4o), Azure AI Search, Cosmos DB, Azure Blob Storage, Azure Cognitive Services Speech, Azure App Service, Azure Container Registry, Retrieval-Augmented Generation (RAG), Server-Sent Events (SSE), Docker, Terraform, GitHub Actions, REST, OpenID Connect (OIDC), Multi-Tenancy

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.

• Fine-tuned and evaluated alternative backbone architectures for the AFFGAN handwriting style encoder to better capture writer-specific features, benchmarking against the original baseline architecture and replacing the legacy VGG19 architecture with a custom Transformer-based encoder.
• Extended the system beyond word-level synthesis to line-level handwriting generation, integrating and adapting an existing line-generation architecture to support full-sentence output.
• Deployed the handwriting generation pipeline as a web application, enabling users to select a writing style, input custom text, and instantly generate a corresponding handwriting image.
• Integrated TrOCR into a teacher–student distillation framework to evaluate and improve readability of generated handwriting, improving OCR accuracy by 20% over the baseline.
• Boosted user-study pick-rates by 40%, demonstrating a significant leap in the perceptual quality and realism of the generated handwriting compared to the baseline model.

Tech: PyTorch, Python, GANs, CNN, ViT, TrOCR, Streamlit, CUDA.

Transmission system operator – network provisioning platform (team lead)

• Designed the saga orchestrator in Python that provisions network fabrics with rollback.
• Split IP, identity, and resource allocation into bounded contexts with their own databases and typed clients.
• Implemented data access using DDD repositories and units of work, with no SQL in the domain code.
• Required integration tests using real PostgreSQL and RabbitMQ containers, and API fuzzing for every service.
• Introduced an agent workflow where a spec starts each change and a human approves every merge.
• Rolled out OpenTelemetry auto-instrumentation through GitOps across eighteen Kubernetes clusters.
• Led incident response for stuck provisioning and recorded every outage as a post-mortem.

Languages and software used: Python, FastAPI, Pydantic v2, SQLAlchemy async, PostgreSQL, RabbitMQ, Celery, pytest, testcontainers, schemathesis, Helm, Flux CD, GitLab CI, OpenTelemetry, pi coding agent.

Software product company – datacenter simulation SaaS (backend development)

• Built the asynchronous simulation pipeline as the second-largest contributor among more than twenty team members.
• Tagged every image by commit SHA and semver in the GitHub Actions deployment workflows.
• Diagnosed the RabbitMQ memory alarm that froze every simulation and fixed the limits in the Helm chart.
• Converted the test suite to function-style pytest, with coverage reported in every pipeline.

Languages and software used: Python, FastAPI, SQLAlchemy, Alembic, Celery, RabbitMQ, Redis, PostgreSQL, S3, EKS, Helm, Keycloak, GitHub Actions, pytest, mypy, ruff, SonarQube.

Internal AI platform – model serving and telemetry

• Built the FastAPI service that deploys open-weight models on GPU tiers as OpenAI-compatible endpoints.
• Upgraded production EKS by two Kubernetes minor versions and fixed zone-bound storage faults.
• Traced every model call, including its token count, cost, and latency, through the OpenTelemetry operator.

Languages and software used: Python, FastAPI, vLLM, LangGraph, pydantic-ai, OpenAI and Anthropic APIs, EKS, Helm, Flux CD, Terraform, GitLab CI, OpenTelemetry, Grafana.

Built Python-based AI microservices and integrations for an AEC/VR Unity-based SaaS app, focusing on LLM/VLM capabilities, retrieval-backed systems, RESTful APIs, containerized deployment, and an automation microservice for the CAD-to-Unity pipeline.

• Developed a custom Hybrid A* based algorithm in C# to simulate hospital scenarios and detect early-stage design conflicts from collision/spatial data and generate structured reports.
• Solved and automated the time-consuming problem of converting CAD files to usable Unity environments with a custom-engineered and real-time pipeline using a ZeroMQ-based communication layer to distribute workloads across multiple processes and achieve real-time performance.
• Built a Dockerized FastAPI pipeline for CAD-to-Unity automation, combining vision-based object matching, image embeddings, and precomputed metadata to automatically map CAD objects to Unity behavior scripts, assign properties, and reduce repeated AI inference calls.
• Created documentation and examples to help technical users understand, configure, and extend the AI automation pipeline.

Conversational AI assistant for cloud infrastructure and security queries

• Designed FastAPI backend with multi-turn conversation handler, token budgeting, and context window management.
• Integrated Amazon Bedrock (Claude 3 Sonnet/Haiku); built RAG pipeline with MongoDB chat history and semantic search.
• Implemented Factory Pattern for modular LLM provider switching; reduced model onboarding effort by 60%.
• Reduced LLM inference cost by 35% through model tiering (Haiku vs Sonnet) and prompt/entity consolidation.

Tech: Python, FastAPI, Amazon Bedrock, MongoDB, Streamlit, Pydantic.

• ▶ Built real-time voice assistant (STT → LLM → TTS pipeline) benchmarking and evaluating multiple STT providers including faster-whisper and Azure Speech. achieved sub-3s latency, Groq API (Llama 3) with multi-turn memory - directly handling edge cases in dictation, names and passcode recognition.

Independent engineering practice focused on custom AI systems, production delivery, and fractional technical leadership.

Selected client engagement: Scalutions

• Role: Serve as fractional CTO and hands-on technical lead, responsible for the architecture and agentic infrastructure behind its managed B2B outbound operation.

• Product: Designed and built OutboundLoop, an agentic SDR operating system for research, qualification, personalized outreach, campaign management, human approvals, measurement, and continuous improvement.

• Scope: Own the full system lifecycle—from business processes and agent behavior to context design, model routing, integrations, evaluation, telemetry, reliability, cost control, and production operations.

• Built and fine-tuned CNN and RNN architectures using transfer learning for real-world classification tasks — core deep learning skills applicable to BMW's multimodal LLM and GenAI vehicle function development.
• Implemented Dropout, Batch Normalisation, and Early Stopping across deep learning experiments; evaluated rigorously using precision, recall, F1-score, and confusion matrices for production-grade reliability.
• Developed an AI-powered attendance management system using LBPH facial recognition, deployed via Flask web interface with real-time SMS notifications — demonstrating end-to-end AI product delivery for real users.
• Collaborated across cross-functional teams to deliver scalable, documented ML pipelines designed for reproducibility — matching BMW's interdisciplinary team and research environment.
• Integrated AI tooling directly into the development workflow from design through to testing, maintaining high velocity without compromising correctness.

• Taught Python, Pandas, data analysis, visualization and Power BI to 250+ students.
• Guided practical projects from data preparation and exploratory analysis through visualization and dashboard creation.
• Explained complex analytical concepts clearly to audiences with different levels of technical experience.

• Built a Python pipeline that converts lectionary web content into structured daily JSON, applying liturgical-calendar rules for accurate weekday and Sunday coverage.
• Shipped a read-only FastAPI REST API with shared Pydantic models and delivered a Kirundi-first web client for browser and mobile use.
• Owned the data layer and backend architecture, collaborating closely on system architecture and interfaces while automating refreshes with GitHub Actions and validating the ETL with pytest.
• Impact: Created a reliable, API-driven source of truth for daily Bible readings in Kirundi, enabling consistent access to previously unstructured content.

End-to-end responsibility for all products – conception, architecture, development, and production operations as the sole developer; also customer discussions, proposals, and marketing.

Underwriting Copilot – AI assistant for industrial insurance (in production with a customer)

• Supports underwriters in analyzing industrial insurance submissions – in production at an industrial insurer.
• Framework-independent RAG architecture with Hybrid Search (BM25 + pgvector) across large, heterogeneous document collections.
• Two-stage evaluation and observability pipeline (code assertions + LLM-as-Judge) that makes answer quality, retrieval accuracy, and citation integrity measurable in a regression-safe way.

Kaiki Menu Analyzer – Data intelligence platform (in production with a customer)

• Automatically collects and analyzes food data from around 25,000 German restaurants.
• Scalable 7-container architecture (FastAPI, partitioned PostgreSQL, Redis/RQ) with LLM-supported extraction of structured data from PDF, HTML, and images.
• Complete CI/CD pipelines (GitHub Actions), production cloud deployment, interactive dashboards (Dash).

Kaiki GEO Atlas – GEO platform (in production with a customer)

• Measures brand visibility across five AI engines (ChatGPT, Gemini, Perplexity, Grok, Claude), each augmented with web search, orchestrated as a DAG workflow pipeline (Dispatcher → Sub-Workflows → Scoring → Report) with fail isolation.
• 6-container deployment (FastAPI, Celery, Redis, PostgreSQL); LLM cost estimation, PDF audit report, rule-based cross-signal insights (no additional LLM effort).

Data Pipeline & Analytics Platform – Competitive analysis in the automotive aftermarket

• Automated data pipeline with gap analysis algorithms and role-based access control; 230+ tests.
• Backend with FastAPI, PostgreSQL, SQLAlchemy.

Product development (actively in development)

BankingGPT – AI assistant for complaint management in cooperative banking

• Security architecture at its core: no AI draft reaches the customer without human approval – the approval decision is stored in auditable code, not in the language model (monotonic: the model may increase the level, but never lower it).
• Genuine agentic components, each with its own boundaries: the model selects tools itself through an MCP server (read-only, allowlisted, capped, fail-safe); sensitive cases are sent via the open A2A protocol (JSON-RPC, Agent Card, message/send/tasks/get; client implemented independently) to a standalone specialist agent (securities/law), which never lowers the review requirement (pinned by tests).
• Evaluation-driven across ten analysis rounds; an independent review and blind tests uncovered a security defect that nine automated runs had missed.
• Voice AI frontend, responding live: covered cases are answered during the conversation, while sensitive cases are escalated before generation; measured response latency < 7 s (local GPU STT/TTS).

Stack & production readiness: Python, pydantic-ai, FastAPI/Celery, PostgreSQL/pgvector, FastMCP, fasta2a, Docker; multi-tenant (physical vector isolation per tenant), encrypted PII, OWASP-LLM-reviewed, 275 tests, CI/CD; provider-portable (Ollama / EU-Cloud-Vertex).

After-Sales Assistant – agentic RAG/GraphRAG assistant based on public OEM manuals (automotive after-sales)

• Genuinely agentic with LangGraph: ReAct agent with four tools and conversation memory – the model decides itself whether to use the manual (RAG, Chroma), a knowledge graph (GraphRAG, Neo4j/Cypher – decoding warning lights), or a workshop/booking service.
• Human-in-the-loop before the irreversible action: before every appointment booking, the graph pauses (interrupt) and obtains the driver's explicit confirmation – the same approval-before-action discipline as in BankingGPT, in a different framework.
• Eval as a CI gate: three-part scorecard (RAGAS grounding + deterministic tool-routing accuracy + DeepEval safety: does the answer mention the warning first in the case of a critical warning?) blocks the pipeline; provider-agnostic (OpenAI/Azure/Anthropic), FastAPI with token streaming.

Stack: Python, LangChain/LangGraph, Chroma, Neo4j, RAGAS/DeepEval, FastAPI, Docker.