Three systems in production, one method in daily industrial use.
Autonomous LLM pipelines, a self-hosted agent stack, and a sub-second knowledge base - each running today, each with the numbers to back it. Here is what the problem was, what I built, and what it does now.
48 seconds, three of the systems below - real terminal recordings, no mockups
Current role - step by step
What I do at Kyocera, in the order I do it.
Kyocera Automotive and Industrial Solutions, Dietzenbach: working student 05-09/2026, full-time since 10/2026 as AI Engineer - Digital Transformation. Most of the work is in-house consulting, then the building; it runs in four steps.
01Map
Interviews and workshops with the people who do the work - quality, customer service, sales, change control - turned into process maps.
02Find
A problem register built from what people said, de-duplicated and ranked with its sources, so management can see what to fix first.
03Solve
Working tools on Microsoft 365, not slides: a Copilot agent that drafts customer-service replies and never sends on its own, an Office Script that flags changes in weekly order files, and Power Automate flows.
04Check
AI tools and platforms checked for EU hosting, the EU AI Act, GDPR and cost, a review of the AI usage policy, and decision papers for management.
The research side of the same question is my master's thesis at KIT, submitted 09/2026: a multi-criteria decision model for AI adoption in manufacturing. The four steps are described in more detail on the method page.
ProblemReading market and news feeds by hand does not scale and does not run at 3am.
BuiltAn autonomous pipeline that ingests market and news data, runs the LLM analysis, and publishes signals to Telegram on a schedule with zero manual steps. The model layer is vendor-neutral - Anthropic, OpenAI, DeepInfra, or a local Ollama model - and a shadow-evaluation gate only promotes a cheaper model once it matches the reference on live data.
deepsignal
docker ps - all services healthy on the server
~95%lower model cost after the shadow-eval gate
hardened by backtesting + a 54-agent adversarial review / self-hosted in Docker / months in production
ProblemOne assistant cannot be an expert at everything, and a cloud assistant does not run on your own hardware or your own terms.
BuiltA self-hosted assistant that runs 24/7 and coordinates a library of specialist agents through a meta-orchestrator, with custom MCP servers, voice in and out, and direct email and calendar tools.
"why isn't the graph loading on my phone?"
berezka-site-optimizer - owns berezka.me end to end
one orchestrator, 12 division directors, 200+ specialist agents - hover any node for its name, one real routed example traced in bronze
PythonClaude APIMCPWhisperDocker
🔎In production
Hybrid RAG knowledge base
ProblemA 4,500-note knowledge base is only useful if the right answer comes back in one second, not by scrolling through files.
BuiltA hybrid retrieval system - dense embeddings (bge-m3) plus sparse full-text search (FTS5) plus a reranker - re-indexed every hour and served by an always-on daemon.
The Obsidian vault that powers the hybrid RAG system above - every dot is a note, every line is a link between two of them. Titles hidden on purpose; the structure is the point. Top 700 nodes shown.
4,540 notes13,024 links7,964 files indexed for retrievalre-indexed hourly
The graph didn't load - probably a network hiccup.
Let's talk
Want to talk about any of this?
Questions about one of these systems, or an idea to compare notes on? Write to me.