AI Solutions Consultant for Procurement | Berlin 🇩🇪
10+ years leading procurement and category management at TeamViewer, Scout24, Foodpanda and Delivery Hero — now engineering the AI systems that will transform the function I know inside out.
I don't just advise on AI transformation. I build the tools myself.
Every project here started from a real problem I encountered running procurement teams: manual triage, supplier compliance gaps, fragmented spend data, slow RFP cycles, and market intelligence that arrives too late. These are my answers — designed by someone who has lived them and built by someone who can now ship them.
AI Integration Bootcamp @ Ironhack · MBA-IT. I ship the tools, not just the slides, every project below is live and demoable.
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 | Links |
|---|---|---|
| 🧠 AI-native procurement OS | TrueSpend. n8n + Claude Sonnet 4.6 + PostgreSQL + React. 17 autonomous workflows, 32-table schema (dbmate-migrated), full P2I lifecycle. Role-gated Operations Board (procurement, IT, controlling, legal, admin). 4-agent compliance onboarding, DocuSign JWT Grant, Grafana, Jira ≥€100k. DB-enforced security: NOSUPERUSER PostgREST role, every status transition through SECURITY DEFINER RPCs — PATCH tickets.status → 403 at the database layer. Agent security: inbound email/invoices run under "the LLM advises, deterministic code decides" — model output is schema-validated against an action allowlist, ticket/PO ids are derived deterministically (never from the model), and a prompt-injection repro proves the guard inert; money RPCs gated fail-closed off the browser token. |
GitHub · Live Demo |
| 🔴 Full-stack AI procurement intelligence platform | SpendLens. 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. | GitHub · Live Demo |
| 🏗 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. | GitHub |
| 📦 AI-native supply-chain OS | SCM MASTER, AI-native supply-chain OS unifying procurement, transit-warehouse flow, and full asset lifecycle. FastAPI + SQLAlchemy 2.0 + Pydantic 2 (SQLite→Postgres), JWT role-gating, 52 test files · CI-gated ≥80% coverage, 5-job CI (lint · Postgres · SAST · CVE-audit · agent-safety). Multi-sourcing core: Product decoupled from ProductSupplier (lead time, MOQ, price, rank) — re-sourcing a line is one FK repoint. Serial-tracked Asset traced end-to-end (RECEIVED → … → DISPOSED) with an unbroken provenance link to its PO line. Contract lifecycle + budget burn, capacity with one-click rebalance, and an autonomous weekly purchasing run (demand-justified, one-PO-per-supplier, approve→place) under the rule "the LLM advises, deterministic code decides" — the confidence score is itself deterministic and audited (factor-by-factor), gating auto-place at ≥0.90 confidence & <€200k — proven by a 29-scenario agent-safety harness that feeds the gate hostile AI advice (unapproved supplier, over-cap spend, prompt injection, poisoned calibration) and asserts it refuses, every time. Real inventory science: a Syntetos–Boylan classifier routes each SKU's demand to the right forecaster (run-rate vs intermittent), now backed by Nixtla statsforecast (Croston/SBA) with conformal prediction-interval safety stock — chosen over a hand-rolled TSB on a walk-forward benchmark (the honest finding: lumpy demand is absorbed by stock, not forecasts); service-level safety stock (z × σ over lead-time buckets) + ABC per-class service levels. Learning layer: rule-based threshold calibration from human approve/reject outcomes, with a LightGBM + SHAP calibrator running in shadow mode beside it — advisory, logged, never deciding (the documented, explainable path from rule to ML; a single tree overfits, so it's a regularised ensemble, and it declines when undertrained). Deployed as two fully-isolated stacks (separate Railway projects + Postgres) — a self-wiring public demo and a forge-locked production (refuses to seed, ship demo accounts, or run on non-persistent storage; the weak-admin refusal is regression-tested); production-hardened with row-locked write guards, a health-checked connection pool, and indexed hot paths so the guards hold under concurrency; each with its own analytics cockpit. Cost-intelligence layer: a clean-sheet should-cost engine (components indexed to commodity markets → a defensible cost floor + target price, so you negotiate from our number, with DRAM/NAND sensitivity), and full per-asset TCO (acquisition + landed + deployment + lifetime OpEx + EOL − recovery) rolling up to a correctly-defined TSCMC % — deterministic engines, the LLM only proposes. Spend analytics slice by year — every euro traces from a received Asset back to its order line, so spend rolls up per calendar year (multi-year history, not one all-time blur). |
GitHub · Live Demo |
| 📈 SCM Power BI Cockpit | AI-adoption consulting case for a non-technical CEO: should a cloud/hosting enterprise invest in AI demand forecasting? Synthetic-data generator → 7 internally-consistent CSVs feed a live, auto-refreshing 7-tab web cockpit (Node + Chart.js — Overview, SC Scorecard, Spend, Inventory, Forecast, Should-Cost margin-lever, TCO; cross-filter, click-to-drill KPIs, per-year spend/forecast slicing (a 5-year period selector where the data has a timeline — spend & backtest — with inventory honestly labelled a live snapshot, not a fake rewind), dynamic reorder alerts, forecast why-it-missed/how-to-fix diagnostics) and a Power BI report on the same live API — DAX measures anchored to SCOR DS, forecast accuracy (WMAPE / Bias / RMSE), should-cost & TCO. Backed by cited market research (Stanford AI Index, McKinsey, chip-geopolitics) and a hype-vs-evidence analysis driving an invest / wait / pilot recommendation, with a phased implementation plan + cost/timeline. | GitHub · Live Dashboard |
| 🔍 Market intelligence sub-agent | HERMES. 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. | GitHub |
| ☠️Supplier due diligence agent | HADES. 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. | GitHub |
| 📊 Marketing Channel Statistical Analysis | Full statistical pipeline for $500K marketing budget allocation across 7 channels. Welch t-tests, Bonferroni + BH-FDR correction, bootstrap CIs, Cohen's d. All 14 CPA pairs significant post-FDR. Executive memo with data-backed reallocation. | GitHub |
| 🧪 LLM Evaluation Framework | LangSmith evaluation lab for procurement compliance Q&A. Custom 20-example dataset, LLM-as-judge correctness + completeness evaluators, A/B model comparison (gpt-4o-mini vs gpt-4o). | GitHub |
Production multi-agent architectures, self-healing infrastructure, and real-time AI applications running live.
| Project/Description | GitHub |
|---|---|
| ⚡ Pantheon OS — Autonomous Trading Orchestrator — 8-agent system live on Hetzner, self-scheduling every 15 minutes. ZEUS orchestrates: Icarus (Hermes signal watcher) → Hades (OFAC/EU sanctions firewall) → Artemis (VIX + macro regime) → Pythia (Kelly-inspired position sizing) → Zeus (Claude Sonnet 4.6 reasoning + ChromaDB KB) → Ares (IBKR bracket orders: entry + 3% SL + 6% TP) → Argus (drawdown kill switch). Apollo runs daily: arXiv ingestion, earnings enrichment, self-improvement loop. Agent seniority system: TRAINEE → DIRECTOR, gated by verified win rate. Kafka event bus. Supabase + Grafana. | GitHub |
| 🤖 Icarus AI — Personal Operating System — JARVIS-style AI OS via Telegram + PWA. 20+ capabilities: voice input (Whisper), multimodal document analysis, Gmail/Calendar/GitHub integration, proactive alerts, expense tracking, LinkedIn posting, live web search. Multi-model routing. Persistent memory via Upstash Redis. ~€8–9/month. | GitHub |
| 🔧 ICARUS Self-Healing System — Icarus diagnoses and repairs its own runtime errors. Catches exceptions → Claude reads broken file + traceback → generates corrected version → commits via GitHub API → Railway redeploys (~90s) → Telegram confirms fix. Escalates if same file fails twice. | GitHub |
Self-hosted reliability and security tooling that keeps the production stack healthy — observe-only guardians, firewall hardening, and automated secret hygiene.
| Project/Description | Repo |
|---|---|
| 🛡 Lookout — Docker Host Guardian — Observe-only watchdog for the production Docker hosts. Samples every container's CPU + memory each minute; on a sustained runaway it applies a reversible CPU cap (the only automatic action) and alerts via Telegram, leaving pause/restart/kill as owner-gated commands. Plus: firewall hardening (ufw + DOCKER-USER conntrack rules that actually block Docker-published ports), short-lived auto-rotated service tokens (no long-lived credentials on disk), a repo secret-scanner that watches all public repos for exposed keys, and a push-based health feed so the ops assistant can answer "are the servers running well?" in natural language. | Private repo |
AI systems built and deployed for real organizations.
| Project/Description | GitHub |
|---|---|
| 📊 Client Dashboard — Internal agency dashboard for monitoring all live client AI systems. Real-time status, deployment health, pipeline metrics across projects. | GitHub |
| 🧙 Agency Wizard — Internal onboarding wizard for deploying full AI automation stacks to clients in a single 3-hour session. Validates every credential live, then provisions into the client's own n8n Cloud instance. | GitHub |
| 🩺 AI Triage System (Metabelly) — Autonomous customer support triage for a Croatian gut health brand. Incoming emails classified by AI (category, priority, language), auto-replies drafted, Calendly links appended, results routed to Slack. n8n + Mistral AI + Gmail API. GDPR-compliant. | GitHub |
📧 Noosphr Email Router — AI email triage for Noosphr's inbox. Claude Haiku classifies and routes to #business, #support, or #spam Slack channels with one-click reply buttons. Runs as systemd service on Hetzner VPS. |
GitHub |
| Project/Description | GitHub |
|---|---|
| 🏥 Kita Connect — Full-stack daycare management platform for German Kitas. ~€0/month, GDPR-compliant, Frankfurt-hosted. Three portals: parents, educators (AI-assisted learning stories via Claude Haiku), management (multi-channel comms, automated registrations). | GitHub |
| 📌 Aushang — Digitization for old-school German orgs (Kitas, Vereine, Kirchengemeinden, Kleingärten) that changes none of their processes: they keep pinning paper to a physical board; one admin photographs it from inside the tool, and members get a private feed, a shared calendar, an ICS subscription, and an email digest. Privacy by construction — the raw photo is OCR'd and PII-redacted locally (Tesseract + Microsoft Presidio + spaCy, fail-closed) before only the redacted text reaches the LLM (Claude, US — never raw images or PII; swappable to an EU model); raw photos and the LLM key never leave the FastAPI worker. "The LLM advises, deterministic code decides" — nothing reaches members without explicit admin confirmation, and all model output is schema-validated. Hardened to a four-layer security model (deny-by-default middleware → server role checks → SECURITY DEFINER RPCs → Postgres RLS + column-level REVOKE on PII), put through multi-agent adversarial security reviews. Next.js 16 + React 19 + Supabase (EU, RLS on every table), a Dockerized Python ML worker, a native Android app (Capacitor), and a one-command self-host wizard. | GitHub Self-host |
| ⚡ Light-weight Transcriber — Drop a YouTube URL or paste any text. Ask Claude anything about it. Answers without downloading the audio — paste a URL or text and ask. | GitHub |
| Project/Description | GitHub |
|---|---|
| 📚 RAG Pipeline — Chunking, embedding, retrieval with metadata filtering. Upstash Vector, OpenAI embeddings, query pipeline with source tracking. | GitHub |
| ⚖️ Relevance Scoring & Rerankers — Advanced RAG over EU AI Act legal text. Vector similarity, metadata filtering, Cohere cross-encoder reranking, before/after position-shift analysis. | GitHub |
| 🤖 LangChain Tool-Use Agent — ReAct-pattern agent with free tool selection across 4 custom tools. | GitHub |
| 🔄 LangGraph Complaint Processor — Deterministic 5-node state machine with human-in-the-loop checkpoints. | GitHub |
| Project/Description | GitHub |
|---|---|
🧠 TrueSpend Workflows (17) — intake_receiver, chat_assistant, board_action, supplier_reply_handler, docusign_sign, docusign_callback, contract_watcher, reorder_trigger, hyperscaler_monitor, supplier_onboarding, invoice_processor, delivery_confirmation, asset_depreciation, llm_consumption, rag_embedder, dispatch_drain, vps_monitor. Production-grade: 120s timeouts, 3× retry, per-signal trace logging. Status transitions call SECURITY DEFINER RPCs — no workflow writes tickets.status directly. |
GitHub |
| 🏗 Procurement Triage Workflows — 6 importable n8n workflows: PR ingestion, tier routing, ERP budget/PO, RFQ/RFP outreach, quote collection, approval handling. | GitHub |
| 📰 arXiv Research Summarizer — n8n + Claude + Notion. POST an arXiv URL → fetch metadata → Claude summary → Notion record. | GitHub |
Procurement & Strategy
Engineering
Data & BI
| Company | Role |
|---|---|
| TeamViewer | Lead Procurement & Category Management |
| Scout24 | Senior Procurement Manager |
| Delivery Hero / FoodPanda | Category Manager |
10+ years in procurement, now building the AI systems I wished existed when I ran the function.