13.07.2026 aktualisiert


nicht verfügbar
info: nicht verfügbar bis 01.02.2027
(Hands-On) Senior/Lead Architect, Golang, LLM/GenAI, VR/AR/MR, Embedded, WebRTC, AI Workshops
Assling, Deutschland
Nur Remote
Dipl. Ing (FH) Elektrotechnik - AutomatisierungstechnikÜber mich
Experienced Project Lead, CTO & Architect with hands-on expertise in IoT, Edge, AI/LLMs, video streaming (WebRTC), VR and backend systems (Go, Quarkus, Node). Skilled in UX/UI, observability, big data, and multi-tenant architectures. Driving innovation, strategy & AI adoption. Leading tech teams
Skills
AI AgentsLLMAugmented RealityCUDANode.jsTensorFlowTransformerStreamingWebRTCVirtual RealityGpu ProgrammingBackendArtificial Intelligence Markup Language (AIML)Mixed RealityGolang
As a Project Lead, Lead Architect, CTO and Innovation Architect, I specialize in designing and implementing (hands-on) complex systems across a variety of domains including embedded, edge computing, backend development, machine learning and mobile. I use mainly Golang, Java, Quarkus, Node.js (NestJS), Python for machine learning and Angular for web development, Android and Ionic, Electron, Wails and Tauri for desktop & mobile applications. I am also working with IoT/Embedded and edge devices, sensors and have a complete hardware lab at my location available.
I have a deep background in video streaming technologies such as WebRTC/gstreamer - in specific PION and LiveKit - and machine learning frameworks including OpenCV, PyTorch, and TensorFlow.
Since 2023 i am working with LLMs and tools like ChatGPT, Claude AI/CLI, various open source LLM models for local solutions, the LangChain/LangGraph tool family, Ollama, vLLM and LM Studio and various vector databases. I am using local LLMs for privacy-focused use cases. I am proficient implementing Agents, different RAG and related solutions. I work with to MCP- and A2A-Agent use cases. I do workshops for companies covering "The Art of using AI Agents in your dev team" for example.
In nearly all my recent projects i introduced "Observability" solutions based on OpenTelemetry (OTel), Prometheus, Grafana, Jaeger, fluent-bit. I often introduce topics like application KPIs, metrics and then help the teams to implement respective code.
My role has often involved acting as a senior technical liaison for both local and offshore development teams, notably establishing successful collaborations with teams in India - including on-site visits in India. My services include designing interfaces for embedded/edge systems, mobile solutions, and multimodal UIs, with a particular emphasis on integrating speech dialog systems and developing NLP and LLM-related solutions for edge devices.
I offer workshops/technical consulting for companies and teams for topics like NATS, LLMs, GenAI, Microservices etc. These are usually short-term projects and are not listed in my history.
- Core SkillsLanguages & Frameworks: Golang, Java, Rust, Typescript, Python, Quarkus, Node.js (Nest.js), Angular, Vue, Android, Ionic, Electron, Tauri, Wails
- Technologies: Docker, Kubernetes, WebRTC, gstreamer, OpenCV, PyTorch, TensorFlow, LLM tools, OpenXR, CloudXR and models.
- System Architecture & Design: Complex system analysis, UX/UI design, big data system architecture, stakeholder requirement elicitation, offshore team management.
- Industry Experience: Embedded and edge computing, machine learning, video streaming, mobile application development, big data and IoT.
- Bridging the gap between technical and non-technical stakeholders to drive project success.
- Continuously exploring emerging technologies to enhance system efficiency.
Substack Articles: https://juergenfey.substack.com/
GitHub: https://github.com/smarttechlabs-projects
Erfahrung
Berufserfahrungmehr als 10 Jahre
Freelancer seit2000
Sprachen
DeutschMutterspracheEnglischverhandlungssicher
Projekthistorie
Commencing with phase 2 now
Defined the architecture and built the proof of concept for a semantic caching layer that reuses prior LLM responses (text/audio and video) to cut latency and inference cost for a German-first, real-time voice-AI agent. If normalized content is detected, the existing result (text, audio, video) is returned instantly without re-generation.
- Semantic cache: Designed a two-stage retrieve-then-rerank pipeline — bi-encoder vector similarity (pgvector) fused with lexical search via Reciprocal Rank Fusion (RRF) for recall, gated by a cross-encoder
reranker for precision. Calibrated similarity thresholds empirically against a human-labeled German benchmark (GermanSTS) rather than guessing them.
- Pluggable model layer: Implemented a generic, OpenAI-compatible endpoint adapter
- Retrieval research: Evaluated RAG, Graph-RAG, and hybrid storage/retrieval strategies; proved that cosine similarity alone cannot separate adversarial near-duplicates.
- Test-data & evaluation platform: Built a 34M+ sentence German evaluation corpus with a normalization/import pipeline, plus a browser-based testbench for interactive cache-quality experiments, and statistical tooling to derive and validate precision/recall thresholds per model.
- Agent & WebRTC streaming integration
Tech: Go, pgvector/PostgreSQL, vLLM, LM Studio, BGE-M3 embeddings, cross-encoder rerankers (bge-reranker-v2-m3), RAG / Graph-RAG / hybrid retrieval, Prometheus, ffmpeg, DGX Spark, Docker.
As Lead Architect, defining and implementing a scalable cloud-based VR/AR/MR streaming platform. Defining and implementing the system architecture for VR/AR/MR client devices, the backend infrastructure and the integration of typical XR platforms like Unity, Unreal etc. Defining compliance and test cases and related test clients. Enable VR/AR/MR clients to connect to the backend infrastructure. Integrate external including 3rd party APIs for Agents and other use cases. Analyzing/driving requirements and implementing the respective backend features. Preparing the setup for a planned Kubernetes-based deployment.
Integrating a complex WebRTC microservice, a efficient rendering pipeline to provide real-time audio, video and data connectivity for a significant number of users. Connecting Unity/Unreal clients with the WebRTC system. Defining and implementing additional microservices to manage users, VR related content, system notifications and security. Setting up a complete observability platform managing logs, traces and KPIs/metrics.
Exploring options for a variety of use cases including the usage of AI Agents, LLM models and NLP solutions.
Developing use cases and adding related documentation and Jira stories. My role combines Business Analyst, PO, Architect and for some services and PoC also backend developer.
Topics:
- Quarkus
- Golang
- WebRTC
- Unity
- Unreal
- OpenXR
- CloudXR
- LLM
- Prediction models
- Docker
- Kubernetes
- Observability
- PostgreSQL
- NATS
- TURN/STUN
- gstreamer
- ffmpeg
- OTel
- GPU scheduling
- CUDA
Designed and built a privacy-focused platform where creative teams (studios, bands, photographers, event crews) store, share and review large media files, from a desktop app (macOS, Windows, Linux) with a self-operated cloud backend. It combines versioned file spaces, optional end-to-end encryption, encrypted USB backups, VideoCalls with a shared whiteboard, team messaging, and browser-based media intake from outside contributors. One account can work across several workspaces, with strict per-workspace data isolation enforced in the database.
- Built a resumable, checksum-verified upload engine (BLAKE3) for multi-gigabyte files, tested against proxy timeouts and interruptions.
- Implemented opt-in end-to-end encryption (X25519 key exchange, recovery phrase) next to server-side per-workspace encryption with key rotation.
- Designed multi-tenancy with workspace memberships. Database constraints guarantee every file stays in its workspace, and an automated test sweep checks every API route for cross-workspace leaks.
- Integrated WebRTC VideoCalls (LiveKit): guest access, a waiting room, co-hosts, an in-call chat, and a shared whiteboard that can use a picture as its background.
- Built browser-based contributor uploads that can be added to a phone's home screen, plus a review and approval workflow.
- Run production on a single hardened server, with scripted deploys, pre-deploy backups, monitoring checks and an operator admin panel.
- Shipped a user guide and website.