1) Software Engineering Approach
- End-to-End Engineering Lifecycle: AI and software systems are approached as production software, not isolated prototypes. The engineering lifecycle covers requirements and acceptance criteria, system and data architecture, implementation, automated verification, integration, deployment, operational validation, and continuous improvement.
- Quality, Release, and Operations: Delivery is supported through static analysis, type checking, unit and regression testing, integration validation, AI/RAG evaluation gates, CI/CD, smoke checks, post-deployment health checks, rollback-oriented release safeguards, auditability, observability, and measurable technical and operational KPIs.
- Lifecycle Coverage: Requirements & acceptance criteria → Architecture → Implementation → Testing & evaluation → Security review → CI/CD → Deployment → Operations → Continuous improvement.
2) AI Automation & Enterprise Delivery
- OfferGenerator (Flask → PowerPoint): implemented, tested, and deployed on a domain network; manual offer creation was reduced from 4–5 hours to ~30 minutes (approx. 85–90%). More information. See demo.
- Tender/Ausschreibung Generator (Flask → PowerPoint): designed for structured generation of BA/MSc announcements; PDF extraction, template-driven PPTX generation, and Azure/OpenAI-based text generation are integrated. Source content is derived from official PEM study pages (see: pem.rwth-aachen.de/go/id/fecr ). Automated export to network locations is supported. More information. See demo.
- PowerPoint Add-in (C#, .NET Framework VSTO/COM): implemented, tested, and deployed on a domain network as an internal productivity tool. GPT-enabled text and image generation, image analysis, and text-to-diagrams/charts workflows are provided; production deployment for staff use has been established. More information. See demo.
- AI-based Local-First Marketing Agent Framework: architected and implemented as a local-first AI marketing operating system for brand workspace analysis, campaign generation, platform-specific post package preparation, media management, approval-gated publishing, simulation, analytics import, learning loops, and guarded connector workflows. The implementation combines a Next.js dashboard, FastAPI backend, Campaign Autopilot services, safe credential handling, scheduler workflows, local and stub LLM modes, YouTube OAuth publishing, TikTok and Instagram dry-run scaffolds, and Paperclip/MiroFish extension points. The verified engineering baseline includes 131 backend tests, Ruff and ESLint checks, a Next.js production build, healthy launcher and API routes, dry-run publish audit jobs, and analytics and scoring checks; final deployment-readiness tests are running before production deployment. More information. See demo.
- AI-based HR Voice Agent Framework (08/2026): conceived, architected, implemented, tested, and operated end to end as a proprietary AI voice screening framework for structured first-round interviews. The system combines a consent-gated candidate portal, real-time STT-LLM-TTS conversation flow, versioned interview rubric, asynchronous post-call analysis, quote-verification gates, deterministic weighted scoring in code, pseudonymized Slack reporting, idempotent webhook processing, dead-letter replay, audit trails, GDPR and EU AI Act-oriented compliance boundaries, 36 offline tests, CI lint and secret-scan gates, measured cost and latency metrics, and live-call validation. The August 26, 2026 milestone documents the completed software baseline and production-tested HR screening workflow. More information. See demo.
3) Production AI & Agent Safety
- Bounded Autonomy: Agentic systems are designed with defense in depth. LLMs may reason, generate, and propose actions, while authorization, business invariants, credential access, and irreversible side effects remain controlled by deterministic application logic and explicit policy boundaries.
- Operational Controls: Engineering controls include least-privilege tool access, approval gates for high-impact actions, idempotent execution, auditable workflows, prompt-injection regression testing, CI/CD security gates, secret scanning, and human review before production-critical changes.
- Governance Positioning: The approach is informed by OWASP, NIST, GDPR, and EU AI Act-oriented engineering principles, without implying formal certification.
4) Enterprise AI (Privacy-First)
- On-Prem RAG (offline / privacy-first): retrieval-augmented generation for internal knowledge workflows; emphasis is placed on governance, operational robustness, data boundaries, deployment constraints, and maintainable architecture in enterprise environments. More information. See demo.
5) Full-Stack & Edge Platform Engineering
- beatandwin.com - Edge-Hosted Full-Stack Web Platform (05/2026): conceptualized, implemented, hosted, and operated since 05/2026 as a bilingual DE/EN electricity-cost and tariff-comparison platform. The system combines Next.js, React, TypeScript, Cloudflare Workers, Cloudflare D1, serverless REST APIs, deterministic calculation logic, SEO/i18n architecture, GitHub Actions CI/CD, 300+ automated tests, post-deployment health checks, rollback-oriented safeguards, and privacy-by-design operation. Visit site. More information.
- Front-End, UX/UI, and Presentation Layer: Layout, typography, spacing, responsive behavior, information architecture, semantic HTML, modular styling, accessible landmarks and labels, optimized assets, and restrained dependencies are treated as first-class engineering concerns. Additional web properties include nosmanosstudios.com and mightysouls.net.
6) Generative AI Media & Commercial Production Framework
- AIMG - Local-First AI Commercial Production Framework (06/2026): conceptualized, engineered, and deployed end to end as a local-first, cloud-ready production framework that turns a text brief and brand references into finished, quality-checked multi-scene commercials. The framework combines a Next.js/React front end, FastAPI/SSE backend, model registry and auto-discovery, ComfyUI/Diffusers/cloud adapters, FLUX/Wan/Kling/Seedance/Veo model routing, LLM-based direction, automated quality gates, MCP-based agent orchestration, production provenance, and server-side cost governance; local rendering is designed for a 12 GB GPU, including 14B video-model operation via fp8 quantization, block swapping, and distilled sampling. More information. See demo.
7) Speech & Emotion Recognition (ASR/SER) Framework
- AI-based Development: Speech & Emotion Recognition (ASR/SER) Framework — Two Models: separate models for Automatic Speech Recognition (ASR) and Speech Emotion Recognition (SER) have been trained via wav2vec2 fine-tuning; both audiofile-based and real-time inference are supported; evaluation is performed via Word Error Rate (WER) and Unweighted Average Recall (UAR).
- HPC Training (RWTH High Performance Computing): training and inference workflows of the ASR/SER models are executed on RWTH HPC (SLURM) to accommodate large-scale datasets and controlled experiments. More information. See demo.
8) Enterprise IT & System Integration (Avantis + BIC)
- Identity & Access: Active Directory and Group Policy Objects (GPO) are administered, including groups/permissions, service accounts, and file server access control.
- Stability & Security: WSUS patching, Sophos Endpoint Security, CommVault backups, standardized rollouts, documentation, and security awareness/testing are applied to support stable operations.
- Site Operations: Multi-site infrastructure work has been delivered at PEM Avantis and the Battery Innovation Cluster (BIC) .
9) XR Learning Technologies (RWTH LearnTech)
- Project Context (04/2022–10/2023): contributions were delivered within the RWTH LearnTech environment (see LearnTech), including the Unity-based xTeach VR project (see xTeach VR).
- Spatial Audio Integration: Google Resonance Audio was integrated and parameterized for improved realism of distance and room perception in immersive learning environments (Unity-based audio pipeline design and tuning).
- Learning Analytics Instrumentation: keyword recognition and interaction events were operationalized via xAPI; statements were collected in a Learning Record Store (LRS) for subsequent analysis.
- Analytics Visualization: Learning analytics were processed and visualized through a React-based dashboard, enabling structured exploration of user behavior and interaction patterns.
- Research Stimuli: Experimental stimuli were created and iteratively optimized to support controlled studies (audio stimulus engineering and quality assurance under research constraints).
- Publication: Teaching and research context, including this project period, is documented in the Osnabrück Teach-R publications (PDF).