Mission  ·  Architecture  ·  Featured systems  ·  Portfolio  ·  Research  ·  Contact

I translate ambiguous business and operational problems into structured, responsible and usable AI, analytics and automation solutions.

Based in Erlangen, Germany, I bring experience across healthcare technology, applied research, enterprise applications and industrial environments. My focus is not a single model or tool; it is the full path from discovery and evaluation to a system people can understand, govern and adopt.

The moving signal represents work progressing through five delivery stages. The architecture is generated locally as SVG, remains fully readable without animation, and disables motion when reduced-motion is requested.

System inputs: uncertain requirements, operational constraints and available evidence.

System outputs: evaluated, reproducible and usable AI or automation workflows.

Control plane: Responsible AI, GDPR awareness and human oversight.

Eight self-hosted FastAPI workflows for document QA, sentiment analysis, news extraction, PR review, research, voice, RAG evaluation and LLM guardrails. The repository includes local-model defaults, bounded inputs, source provenance, API authentication, tests and CI.

Stack: Python, FastAPI, Ollama, Docker | Status: Tested reference implementation

An auditable control plane that converts constrained product briefs into typed contracts and security-gated execution results. It demonstrates async orchestration, explicit state transitions, PostgreSQL persistence, JWT authentication, migrations, CI and Prometheus/Grafana telemetry.

Stack: Python, FastAPI, PostgreSQL, Prometheus | Status: Production-oriented architecture

A local secret-scanning CLI and browser tool that redacts detected values, supports configurable rules and exports SARIF for GitHub code scanning. It keeps source content local and documents the boundary between screening signals and confirmed findings.

Stack: TypeScript, SARIF, GitHub Actions, DevSecOps | Status: Working developer tool

A browser-local application that parses CSV or text-based PDF statements and estimates recurring monthly or annual payments. Financial files remain on the device, and the detector's assumptions and false-positive risks are documented.

Stack: TypeScript, Next.js, CSV/PDF processing, local-first design | Status: Tested local-first application

The portfolio is grouped by engineering signal rather than chronology. Public previews use fictional data unless a repository explicitly documents configured integrations.

The current references below are publicly hosted on GitHub Pages and do not require my local computer to be running. Private APIs and AI credentials are not exposed. Older Cloudflare mirrors are historical snapshots, not the current security-patched builds.

• Applied AI Systems Portfolio - self-hosted AI services with retrieval, evaluation, guardrails and local inference.
• ForgeAI Production Engine - typed multi-agent orchestration with fail-closed QA, persistence and observability.
• Enterprise Hybrid RAG Workbench - 4-mode hybrid retrieval (SQLite FTS5 BM25, 384-d ONNX embeddings, Reciprocal Rank Fusion k=60 and Cross-Encoder reranking), pre-retrieval RBAC, freshness gating, passage-level NLI claim verification and SHA-256 digest-bound HITL review across 50 benchmark cases (top-5 retrieval coverage 1.0000 vs 0.6471 lexical-only, MRR@10 0.9853, NDCG@10 0.9891, passage claim support 0.9868 and 0.0000 RBAC leakage); local CPU ONNX execution verified, external cloud LLM connectors unconfigured.
• Enterprise Hybrid Data Copilot - decision-support copilot bridging 4 relational SQLite KPI tables and 16 operational SOP passages with pre-query data-freshness and null-ratio gates, AST-validated read-only parameterized SQL, deterministic numerical grounding verification and a SHA-256 digest-bound HITL approval state machine (DRAFT -> PENDING_APPROVAL -> APPROVED / REJECTED) across a 40-query Build-vs-Adopt benchmark (1.0000 SQL execution rate, 1.0000 numerical parity vs 0.5100 ungoverned Prompt-to-SQL, 1.0000 stale-partition gate selectivity and 83.24% token reduction); local SQLite/FastAPI execution verified, external cloud warehouses unconfigured.

• Boston Housing Research Workbench - GPLv3 adaptation with upstream history preserved, 11 named inputs excluding B/LSTAT, independent calibration/test splits, actual browser ONNX inference, local scenarios/CSV and held-out evidence. Thirty Python tests, eight browser cases, standalone installed packages, Linux/Windows/container gates and real data/training reproduction pass. Historical tract-median MAE $2,000 versus $5,882 median baseline; not individual/current housing valuation, lending or allocation. Exclusion is not a fairness guarantee; OpenML's Public license label is nonstandard, with no stronger commercial clearance asserted.
• Airbnb Price Workbench - original MIT implementation with separately licensed NYC2019 data, host-disjoint training/evaluation, calibrated uncertainty and actual browser ONNX inference; test MAE $48.80 versus $52.66 borough/room baseline, 18 Python tests, 16 browser cases, 30 sklearn/browser parity scenarios and independent data/training reproduction verified. Historical benchmark, not current rental quotes or bookings; no code/model/data imported from the unlicensed gallery reference.
• Urban Change Detection & Change Captioning - Master's thesis adapting the dual-head SegFormer MiT-B1 and Transformer MCI model on LEVIR-MCI (10,077 original and 15,115 augmented pairs) with bidirectional temporal-reversal augmentation for removal/deconstruction cases, split-family provenance auditing (5,038 reversed pairs mapped 1-to-1 with 0 cross-split leakage), directional consistency evaluation (0.9767 mask symmetry mIoU and 1.0000 caption polarity inversion on the checksum-verified test slice), auditable 4-panel visual report cards with bounding-box coordinates, CLI/Streamlit/FastAPI workflows, 35 pytest checks and SHA-256 verified model weights and augmented dataset; CPU inference and numerical equivalence verified, full-dataset benchmark reproduction and mixed upstream licensing (MIT Change-Agent/COCO-caption plus NVIDIA non-commercial SegFormer terms) documented.
• AI Werewolf Arena - multi-agent social deduction with episodic memory, guardrails, decision tracing and an ELO evaluation harness.
• NeuroDrive - browser-based driving simulation with a neural network and genetic algorithm.

• AI Receptionist - signed Twilio routes, FAQ lookup and Supabase logging; live integrations remain unverified.
• WhatsApp Business Assistant - verified Meta webhooks and intent routing; the public UI is a simulator, not a live messaging integration.
• AI Chief of Staff - Gmail and Calendar briefing prototype; durable Google OAuth is not implemented and scheduled delivery requires configuration.
• Voice Note Idea Catcher - browser recording, optional Whisper transcription and structured action extraction.
• AI Blog Publisher - attributed BlogBoard derivative with private drafts, digest-bound approval and sanitized static articles; tested reference implementation with live AI/search/R2 integrations unverified.
• TravelBrain - attributed GPLv3 trip workbench for Paris, Lisbon and Kyoto with source-linked itineraries, illustrative budgets, browser/private SQLite saves and exports; cross-runtime planning, browser workflows and live Open-Meteo weather verified, Groq inference unverified.

• Security Checker - local credential screening with redaction, SARIF output and GitHub Actions support.
• Subscription Auditor - private, browser-side recurring-payment analysis for CSV and PDF statements.
• Money Dashboard - protected income records and integer-cent revenue aggregation with Prisma-backed SQLite; single-admin demo with fictional sample data.
• GitPulse Audience Workspace - original implementation inspired by GitPulse's documented concept, with atomic SQLite snapshots, append-only observations and a typed dashboard; public REST sync verified, hosted private API unverified.
• Doclify - attributed AGPLv3 documentation CLI with bounded source discovery, incremental caching and digest-bound reviewed publication; offline workflows, packaging and cross-platform CI verified, live Groq inference unverified.
• Market Insight - attributed GPLv3 financial research workbench with bounded Yahoo Finance retrieval, source-dated tables and private SQLite reports; complete sample workflow, Linux/Windows CI and container checks verified, paid AI inference unverified.
• Software Intake Gatekeeper - read-only intake checks for third-party apps and binaries: CycloneDX SBOM from Go build info, Rust cargo auditable data, npm/ASAR, NuGet, Python metadata and Maven pom.properties (including nested Spring Boot/WAR archives), opt-in OSV vulnerability lookup, Electron fuse and webPreferences rules, Authenticode signer identity, SARIF and an expiring-waiver policy gate, plus update checks that gate on new vulnerabilities, weakened fuses, string-literal endpoints, native modules, removed signatures and a changed publisher; v0.3.0 parsers are checked against rust-audit-info on Linux/Windows/macOS, Maven Central and PyPI artifacts (Log4Shell in a nested JAR) and Windows Get-AuthenticodeSignature (real-world case study); published on PyPI and the GitHub Marketplace with provenance-attested artifacts; findings are version-level evidence rather than reachability.
• Privacy Flow Inspector - offline privacy gate for recorded web traffic (HAR): test-persona matching including hashed, Base64 and gzip-compressed forms, WebSocket messages (Socket.IO/SignalR framing), checksum-verified IBAN, Steuer-ID and KVNR detectors, health-data fields, first/third-party data-flow map with third-party eTLD+1 inventory and opt-in connection gating, approved processors and expiring waivers, SARIF and a self-checked redacted HAR; ships a Playwright Test (Node) fixture that records WebSocket messages, and CI requires identical findings from Python and Node recordings with real Chromium; published on PyPI and the GitHub Marketplace; technical evidence for privacy reviews, not legal advice.
• Portfolio Maintenance Controller - bounded noon Europe/Berlin maintenance across seven gallery references: real unit/integration execution, behavioral red/green gates, pinned offline Qwen CPU proposals, separate review contexts, measured local workflows and durable source-SHA reports. Owner-authorized automatic merges require every independent check, unchanged base and narrow scope; protected main branches enforce these checks even for administrators. Verified main CI gates deployment, with bounded reconciliation for missed events. No routine manual approval/start is required. Serial journeys are not concurrent users; automated model review is not independent human pentest certification. Old fixed counts and dialogue remain explicitly labeled legacy fixtures.

The Master's thesis repository and artifacts are publicly available; additional academic and enterprise studies are presented at a sanitized capability level without disclosing internal systems, datasets or business details.

Languages: English C1; German B1 progressing toward B2.

Current focus: Enterprise RAG, AI agents, workflow automation, MLOps and responsible adoption.

Status: Open to full-time Data and AI opportunities.

• Interactive portfolio
• LinkedIn
• GitHub repositories

_{Systems Console is generated locally with Python. It uses no remote image service, font, analytics tracker, client-side JavaScript or profile-statistics server. Previous layouts remain available in Option 1 and Option 2.}