Representative Work Types
32 projects — click for details
Mobile recorder web app (https://the-implementers.de/Recorder/) with topic-based AI pipeline: speech to structured text to automatic routing to EHR/notes/project DB. Versatile: letters, medical histories, team meetings, site surveys, handovers.
Public GitHub project (MIT) for detecting ghost hangs in Ollama LLM inference. 4 collector services, SQLite backend, REST API, dashboard. Real impact: OCR pipeline 97.8% success rate (vs 75-80% naive), hang recovery 140x faster. https://github.com/tiubde-michael/ollama-stallwatch
Central routing engine for 4 heterogeneous GPU nodes and 8+ model endpoints. Capability-based routing, tenant priorities, preemption. 130,000+ routed requests in 14 days, zero-downtime shadow-mode migration.
OpenAI-compatible /v1 gateway (chat/embeddings/rerank) with token scopes and live rate limits, serving an external customer K8s cluster. 338,000 requests and 463M input tokens in 14 days.
Brought an NVIDIA GB10 (ARM64) into production as an LLM server: diagnosed vLLM instability, migrated to llama.cpp with context split. ~193 tok/s at 0 errors, 429 backpressure strategy.
Lead architect of an on-premise RAG knowledge platform for a COBOL codebase of several million LOC. 17,279 chunks with dual embeddings, 3,697 graph edges, hybrid RRF retrieval across 5 paths with mandatory citation. Recall@10 raised from 61 % to over 83 %; model benchmark across 7 candidates as a hardware decision basis.
IMAP automation with rule engine, full-text archive and attachment routing into the OCR pipeline. Second RAG system with 24,500+ embedded chunks in pgvector for an internal assistant.
Built and operate a development environment in which around two dozen Claude Code sessions work in parallel across several machines, each with its own workspace and identity. Coordination via a purpose-built ticket board on PostgreSQL: priorities, due dates, responsibility as a dedicated field, blocking, conditional claim; agents commission and review each other through sub-tickets. Shift operation with a watcher, versioned documentation and playbooks from Git, credentials exclusively via a broker. Codex as an independent reviewer.
Control station for running many Claude Code and Codex CLI sessions at once: nine windows with around fifty sessions, built with one click and laid out automatically per workplace. Simultaneous prompting of multiple sessions with measured input readiness, automated command sequences, detection and reporting of open agent questions. Around 6,400 lines of Python, standard library only, 20 tables, 249 regression tests; replaces 13 shell scripts, monitoring load cut from 200 % to 4 % CPU.
AI-assisted reverse engineering and replacement of an 18-month project plan in Excel with around a dozen VBA macros for timeline, bars and predecessor arrows (around 130 work packages in 12 phases), implemented with Claude Code and reviewed with Codex. Key finding: Excel was carrying two roles at once — data storage and rendering. New architecture: CSV as source of truth, generated SQLite, static single-file HTML — 2,733 lines of Python and HTML/JS with no external dependencies, all four predecessor types (FS, SS, FF, SF), automated consistency check. Two silent error classes of the legacy system eliminated by construction.
Operate Claude Code additionally against locally hosted open-weight models (Qwen on llama.cpp) through a translator between the Anthropic and OpenAI interfaces — agentic coding without a cloud connection. Found and fixed two different causes behind the same grammar-enforcement error message; production run with 40 tools and around 87,000 tokens. Model choice by measurement: 144 agent runs with hidden checks — Qwen3.8-27B solves 72 of 72, Qwen3.6-35B-A3B 64 of 72; routing by task derived from it.
TI-Flows Stack: Self-hosted workflow automation replacing n8n. 20+ workflows, 30 frontend pages, 7-layer security, LLM pipeline, depot analysis. ~24,000 LOC Python/React.
Billing system for medical providers with ZUGFeRD-compliant invoice generation and OpenEMR integration.
On-prem OCR pipeline replacing Azure: 10,871 pages, 97.8% success, extraction 33%→86% vs Azure. 5-layer multi-LLM stack (PaddleOCR, qwen-vl, qwen3.5, mistral-nemo). Incl. bank matching and DATEV export.
Automated HRV analysis from long-term ECG: R-peak detection, artifact correction, PNS/SNS trends, 3-page PDF report. Docker.
Clinical documentation system: OPS tracking, traffic light system, bed occupancy, audit trail. FastAPI + React + PostgreSQL.
Clinical decision support system: automated pain questionnaire evaluation with 20-factor traffic-light scoring and AI-powered admission justification.
AI pipeline for automated medical documentation from physician-patient conversations. Self-hosted LLM, speaker diarization, OpenEMR integration.
Supplier development for forged parts with turning/milling incl. changes, launches and claims management.
Transfer of 150 serial production suspension parts due to short-notice plant closure.
Process optimization in internal sales of a Swiss machinery manufacturer.
Ongoing support for a private practice: electronic anaesthesia record with monitor integration, OpenEMR, GOÄ billing, end-to-end telemedicine workflow from GDPR consent to billing, document management.
Start-up consulting and coaching for a private pain therapy & TCM practice.
Interim plant management and stabilization of a Hungarian production facility until handover.
Supplier development DAG BR238 door trim incl. sampling, PPAP/EMPB and production launch.
Interim COO & plant manager at aluminum-processing automotive supplier. Expansion of Slovakia site.
Cost reduction (process, material, labor, suppliers) in headliner production.
Program management interior door trim DAG C292 incl. industrialization in the USA.
Consulting project team in Turkey for sun visor development and first article inspection.
Cost reduction for interior door trim DAG C218.
Transfer of injection molding tools to new suppliers in running production.
Program management BMW sun visors — new development and production transfer.