Projects

Short for Agentic LLM Orchestration & Integration Service. It handles recurring work: mail triage, a content pipeline, onboarding support and editorial review.

Built with Python, LangGraph, LangMem and FastAPI, deployed on a self-operated K3s cluster. Each persona has its own role description, tool set and memory.

The operational side is designed alongside the functional one. What an agent may do on its own, what needs human approval and how much it may spend is defined up front. Every run can be reconstructed afterwards.

Key Achievements:

• Seventeen personas on one runtime, five of them carrying the recurring work
• Two independent review agents, with a deterministic moderator between them
• Long-term memory via LangMem and pgvector, session state in PostgreSQL
• Full observability in Langfuse: token budgets, cost dashboards, alerting
• Destructive actions held in quarantine for human approval
• Rate limiting and a circuit breaker against runaway loops

The assistant on this site answers questions about my background, skills and projects. Inference runs on hardware I operate myself.

Gemma 4 12B is served by Ollama, with LiteLLM in front as an OpenAI-compatible router. Embeddings come from Text Embeddings Inference and are matched against Qdrant. The Flask API orchestrates intent classification, hybrid retrieval, generation and validation against the retrieved content.

A publicly reachable language model needs the controls around it: prompt injection detection, rate limiting, CSRF protection, and answers derived only from indexed site content.

Key Achievements:

• Runs on a single RTX 3060, on hardware I operate myself
• No data sent to external AI services
• Hybrid search with reciprocal rank fusion and cross-encoder reranking
• Answers validated against the retrieved sources before they are returned
• Content re-indexed automatically after each deployment of this site

A security assessment of the RAG chatbot running on this site, carried out in June 2026 against the live production system.

Method: OWASP LLM Top 10:2025 plus current jailbreak techniques. A white-box audit of the codebase, combined with live testing from inside and outside the network.

The language model layer held. What required action sat in the infrastructure around the model, documented as a prioritised remediation plan with file references and a rotation runbook.

Key Achievements:

• Complete OWASP LLM Top 10:2025 coverage matrix
• White-box audit of the codebase, black-box live testing against production
• The LLM layer withstood every live attack
• The findings that mattered concerned the infrastructure around the model
• Documented as a prioritised remediation plan with a rotation runbook

Integration of the Sibros Deep Connected Platform into Pierer Mobility motorcycles - the group behind KTM, Husqvarna Motorcycles and GasGas - including the KTM V-Twin Series, Super Adventure and Super Duke.

The work centred on connected vehicle capabilities over secure vehicle-to-cloud communication.

Key Achievements:

• Over-the-air updates for software management in the field
• Remote diagnostics for dealers and customers
• Android Automotive OS integration for the global rollout

Twenty years at SPAR ICS, from IT hotline and field support through POS technology management and enterprise architecture to senior product lead. Two projects stand for that period.

eSTORE was the cash register platform: Linux-based, powered by Java, built for reliable operation across regions. A CI, CT and CD framework made frequent releases into live stores manageable.

The Digital Leadership Store brought self-checkout, electronic shelf labels and a snack pre-ordering app together in one store, under real operating conditions.

Key Achievements:

• eSTORE: Linux and Java cash register platform deployed across multiple regions
• Continuous integration, continuous testing and automated deployment around it
• Higher reliability and faster delivery through automated pipelines
• Digital Leadership Store: self-checkout, electronic shelf labels and a snack pre-ordering app