user avatar

About

M.Sc. in Chemistry & Plasma Physics. 16 years of progressive IT: embedded IoT QA → full-stack web → CTO-level ownership of a 70+ node bare-metal cloud and its AI/ML serving stack. Last 5 years: designing CPU-constrained computer-vision and VLM inference chains for edge and cloud. Still the escalation point for the company's smart-home integrations.

Last 5 years dedicated to AI / MLOps: building and deploying computer-vision and VLM inference chains on CPU-constrained edge hardware and cloud, with full CI/CD, observability, and DevSecOps ownership across a 70+ node bare-metal fleet (HashiCorp Nomad / Consul, Zabbix, self-hosted GitLab).

Strong at the intersection of systems thinking from the sciences, hands-on Linux/cloud operations, and applied ML – from a 40 MHz sensor node to a multi-tenant cloud platform.

  My latest projects focus on DevOps and MLOps pipelines and monitoring of cloud applications (online documentation, support ticket platforms, marketing system, SysAdmin dashboards, etc.) as well as backends (search engines, computer vision prediction APIs, provisioning tools, service meshes, etc.) - from the developers PC through source control and onto the internet - agile, observable and cloud-native.

Skills











Experience

Aug2026

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
present

A FastMCP server over the camera's MQTT control API that turns 700+ raw topics into six clean, capability-based tools any LLM agent can call — with state-confidence tracking and credential-safe design, so agents drive the hardware without hallucinated commands. Shipped once; available to every MCP-compatible harness a customer already uses.

Jul2026

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
present

A VLM is a GPU-shaped tool; I had a CPU server and a hard latency budget. Fine-tuned on a custom security-scene dataset (detected by my YOLO detector to prime regions of interest), trained multiple variants to cut token-generation latency, enforced-JSON output, exported to OpenVINO, and served thread-safely inside the Nomad-managed inference job with a shared-volume local model registry. Multimodal understanding on hardware that should not be able to do it, within the latency target.

Nov2025

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
present

The cloud AI pipeline ran distributed across multiple GPU servers, each with its own model, its own crash modes, and no stability under real load. I rebuilt it as one serving stack on a CPU-only server: three resident model servers (YOLOv10 / LiteRT, Fast-Plate-OCR, fine-tuned SmolVLM) loaded once behind a Redis job queue with sorted-set batching and thread-safe worker loops; a watchdog validates each job (JSON schema) and drives a capped Ray actor pool with round-robin dispatch and fleet-wide structured logging; a connector layer writes results to the customer-facing DB, triggers per-tenant notifications, and routes to the least-loaded node. Six months of model optimization (OpenVINO / LiteRT), memory budgeting, and fault-tolerance work drove the failure modes out. Deployed as a Nomad job with canary rollout, bounded retries, strict per-task CPU/memory limits, and a dedicated Zabbix dashboard — a broken model can now take down one allocation, not the fleet. 10k+ detection tasks/min in production.

Feb2021

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
2 yr

Operate a 70+ bare-metal and VPS fleet (Linux, CentOS / Ubuntu) serving internal tooling and customer-facing web services. Orchestrated all workloads on HashiCorp Nomad + Consul: self-healing, auto-scaling, one-PR onboarding for new teams. Standardised CI/CD on self-hosted GitLab: OSS source → Docker container → Nomad deploy → Zabbix / Grafana observability. DevSecOps: CVE response, security monitoring, access-control hardening across every customer-facing service.

Feb2018

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
3 yr

Deployments across a growing fleet were manual, inconsistent, and untracked. I honestly evaluated Ansible / Salt / Kubernetes / Nomad and chose Nomad for its lean footprint: an existing Dockerfile re-parses into a job file and you inherit the orchestration. Today the platform spans 6+ namespaces, 6+ node pools, 10+ datacenters, each isolated with least-privilege Sentinel ACLs, mTLS (cfssl CA), canary + health-gated + auto_revert deploys, Consul service discovery driving nginx upstreams, and Consul-Template for cross-datacenter config. New-team onboarding went from a multi-day manual process to a single job-file PR.

Mar2016

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
2 yrs

Designed and implemented webhook + MQTT broker interfaces directly in camera firmware, making INSTAR drop-in compatible with Home Assistant, openHAB, ioBroker, and Node-RED. To date the only surveillance-camera manufacturer exposing full control over MQTT — and the MCP server that later exposed those same topics as AI-agent tools. QA engineer, Tier-2 support, and first escalation contact for all smart-home integration issues – a role I still hold.

Mar2015

INSTAR Deutschland GmbH (Waletech) Shenzhen, China
Chief Technology Officer
3 yrs

Built the customer knowledge base (Gatsby / Elasticsearch), community forum (Discourse), and SLA-tracked ticketing (osTicket) – all three deployed from zero to production. First- and second-tier technical support; on-site production QA; cross-platform integration testing.

Apr2012

INSTAR Deutschland GmbH Guangzhou, China
General Manager
3 yrs

Leading a team in the development and maintenance of a customer online knowledge base, based Node.js and React.js on a CentOS and Elasticsearch backend. Integration of this online help platform into all in-house software products. As well as setting up internal training programs for the customer service team. New product development and quality assurance in cooperation with the production line.

Mar2011

INSTAR Deutschland GmbH Guangzhou, China
Support Technician
1 yrs

First and second tier technical support for end customers and on-site quality assurance for the production cycle. Creation of print and online documentation for customers and trainings material for new employees. Enabling the system integration into third-party software and hardware products.

Oct2004

University Cologne, Germany
Laboratory Technician
5 yrs

Mass spectrometry and Terahertz spectroscopy for the detection of iron compounds in high energy and low pressure plasmas (Ms.Sc.) for the department of physics and astro science. Combined with a teaching position for laboratory courses for the department of inorganic chemistry.

Mar2002

University Wuppertal, Germany
Scientific Assistant
4 semester

Development of new spectroscopic methods in atmospheric analysis for the department of physical chemistry.

Apr2001

University Wuppertal, Germany
Scientific Assistant
2 semester

Teaching position for laboratory courses in inorganic chemistry and instrumental analysis.

Latest repositories on Github