Procurement Leader → AI Engineer | Berlin 🇩🇪
10+ years leading procurement and category management at TeamViewer, Scout24, Foodpanda and Delivery Hero — now building AI systems for the function I know inside out.
I learned to build the tools myself rather than only advising on them.
Every project here started with a real problem I hit running procurement teams: manual triage, supplier compliance gaps, fragmented spend data, slow RFP cycles, and market intelligence that arrives too late. These are my attempts to solve them — designed by someone who has lived the problems and learned to ship the code.
Completed the AI Integration Bootcamp @ Ironhack Berlin.
AI-powered tools built from 10+ years of hands-on procurement experience — targeting the exact pain points category managers, CPOs, and procurement ops teams face daily.
| Project | Description | GitHub |
|---|---|---|
| 🔴 SpendLens | Full-stack AI procurement intelligence platform. React 18 SPA + FastAPI. 5-stage AI pipeline: column mapping → cleanup → vendor classification → compliance flagging → supplier intelligence. 7 screens: Dashboard, Deep Dive, Compliance Scorecard, CLM, Icarus AI, Supplier DD, Category Strategy. Deployed live on Railway. | Link |
| 🏗 Triage Agent | Autonomous agent replacing manual PR triage. 5-tier value routing, supplier NDA/DPA/MSA compliance check via RAG, RFQ/RFP generation, multi-supplier outreach, evaluation matrix, award recommendation. 6 importable n8n workflows. | Link |
| 🔍 Hermes | Market intelligence sub-agent on Railway. Crawls 590+ suppliers across 17 categories via 5 crawlers (RSS, EDGAR, Tavily, Jobs, Earnings). Signals classified by Claude Haiku with delta tracking. Semantic RAG via Upstash Vector. Powers SpendLens Icarus AI. | Link |
| ☠️ Hades | Supplier due diligence agent. POST a company name, get a full risk report in under 2 minutes. 6 parallel LangGraph nodes: OFAC/UN sanctions, NorthData registry, LkSG/CSDDD signals, ESG, news sentiment, Hermes intel. Risk score 1–10 + Approve/Block recommendation. Integrated into SpendLens. | Link |
Multi-agent architectures, self-healing infrastructure, and real-time AI applications — deployed and self-hosted.
| Project/Description | GitHub |
|---|---|
| 🤖 Icarus AI — Personal Operating System — JARVIS-style AI OS via Telegram + PWA. 20+ capabilities: voice input (Whisper), multimodal document analysis (invoices, contracts, whiteboards), Gmail/Calendar/GitHub integration, proactive email alerts, expense tracking, LinkedIn posting, live web search, Google Maps. Multi-model routing (Haiku for simple, Sonnet for complex). Persistent memory via Upstash Redis. ~€8–9/month. | Link |
| 🔧 ICARUS Self-Healing System — Icarus diagnoses and repairs its own runtime errors without human intervention. Catches exceptions → Claude reads broken file + traceback → generates corrected version → commits via GitHub API → Railway redeploys (~90s) → Telegram confirms fix. Escalates to manual review if the same file fails twice. | Link |
| ⚡ Pantheon OS — Multi-Agent Orchestration System — an 8-agent pipeline under a single orchestrator, deployed on Hetzner, self-scheduling every 15 minutes. Engineering showcase, paper trading only — no real money at risk. The point is the orchestration: one owner of control flow, agents that never call each other, typed contracts, circuit breakers, a watchdog, and kill-switches at every stage. ZEUS orchestrates: Icarus (Hermes signal watcher, 590+ suppliers) → Hades (OFAC/EU sanctions/ESG compliance firewall) → Artemis (VIX + macro regime) → Pythia (Kelly-inspired sizing from historical win rates) → Zeus (Claude Sonnet 4.6 LLM reasoning + ChromaDB KB + ticker history) → Ares (IBKR bracket orders: entry + 3% SL + 6% TP) → Argus (drawdown kill switch + portfolio monitor). Apollo runs daily: arXiv ingestion, earnings enrichment, self-improvement loop. Agent seniority system gates live trading — agents start at Trainee L1 and earn their way up four tiers (Trainee → Junior → Intermediate → Senior, 10 levels each) on verified winning trades; real money unlocks only at the Senior tier and only after a human arms it. Grafana dashboard tracks equity, drawdown, agent health, and Anthropic token spend. Supabase persistence. | Link |
AI systems built against real-world client briefs — real problems, deployed infrastructure.
| Project/Description | GitHub |
|---|---|
| 📊 Client Dashboard — Internal agency dashboard for monitoring all live client AI systems in one place. Real-time status, deployment health, and pipeline metrics across projects. | Link |
| 🩺 AI Triage System — Autonomous customer support triage for a Croatian gut health brand (Metabelly). Incoming emails classified by AI agent (category, priority, language), auto-replies drafted for FAQ and medical deflections, Calendly booking links appended, results routed to Slack. Built on FastAPI + PostgreSQL + Gmail API. GDPR-compliant: content encrypted at rest, audit log append-only, no PII in summaries. | Link |
End-to-end products combining AI reasoning, workflow automation, and real-world integrations.
| Project/Description | GitHub |
|---|---|
| 🏥 Kita Connect — Full-stack daycare management platform for German Kitas. ~€0/month operational cost, GDPR-compliant, Frankfurt-hosted. Three portals: parents (development docs, messaging, multi-language DE/EN/RU), educators (Sismik/Seldak/Perik observation tracking, AI-assisted learning stories via Claude Haiku), management (multi-channel comms — in-app, email, Telegram, SMS — automated registration workflows). | Link |
| 🎙 PodcastIQ — Automated Podcast Studio — Any PDF, URL, YouTube link, or plain text → polished two-host MP3 episode in under 60 seconds. 4 podcast styles, dual LLM (Claude Sonnet / GPT-4o), ElevenLabs premium voices + edge-tts fallback, 3-layer humanizer, emotional prosody, content safety guard, parallel async TTS synthesis. ~$0.03–0.05 per episode. Ironhack Week 2 Capstone. | Link |
🎓 Bootcamp coursework (Ironhack) — retrieval-augmented generation, from basic chunking to reranking and semantic search.
| Project/Description | GitHub |
|---|---|
| 📚 RAG Pipeline — Chunking, embedding, retrieval with metadata filtering. Upstash Vector store, OpenAI embeddings, query pipeline with source tracking. | Link |
| ⚖️ Relevance Scoring & Rerankers — Advanced RAG over EU AI Act legal text + podcast transcripts. Vector similarity retrieval, metadata filtering, Cohere cross-encoder reranking, LLM relevance scoring, before/after position-shift analysis. | Link |
| 🔌 MCP Integration — LangChain tool-use agent connected to a local filesystem via Model Context Protocol. Demonstrates standardised tool discovery and direct-API comparison. | Link |
🎓 Bootcamp coursework (Ironhack) — structured and unstructured agent workflows, from free tool selection to deterministic state machines.
| Project/Description | GitHub |
|---|---|
| 🤖 LangChain Tool-Use Agent — ReAct-pattern agent with free tool selection across 4 custom tools. Agent decides which tools to call and in what order without a fixed script. | Link |
| 🔄 LangGraph Complaint Processor — Deterministic 5-node state machine: intake → validate → investigate → resolve → close. Explicit edges, conditional transitions, human-in-the-loop checkpoints. | Link |
n8n workflows for procurement, research, and operational automation — the procurement set powers the Triage Agent above; the rest are 🎓 Ironhack coursework.
| Project/Description | GitHub |
|---|---|
| 🏗 Procurement Triage Workflows — 6 fully importable n8n workflows for the autonomous procurement agent: PR ingestion, tier routing & SMTP notifications, ERP integration (budget check + PO creation), RFQ/RFP supplier outreach with daily reminders, quote collection, approval response handling. | Link |
| 📰 arXiv Research Summarizer — n8n + Claude + Notion pipeline. POST an arXiv URL → extract paper ID → fetch metadata via arXiv API → Claude summarises abstract → stores structured record in Notion research database. | Link |
| ⚙️ Error Handling & Scheduled Workflows — n8n patterns for production resilience: HTTP retry logic (3×, 5s delay), error output branching, daily scheduled runs with idempotency guard via date-keyed IF node. | Link |
| 🔍 n8n Node Reference — Systematic study of 20 node types across 3 workflows. Parameters, JSON input/output, key transformations, debugging tips. | Link |
Procurement & AI Transformation
Engineering
| Company | Role |
|---|---|
| TeamViewer | Lead Procurement & Category Management |
| Scout24 | Senior Procurement Manager |
| Delivery Hero / FoodPanda | Category Manager |
10+ years of procurement expertise. Now building the AI systems that will transform how procurement teams operate.