Work
I design and ship agentic AI systems for German industry — lead engines built on live CRM data, production logistics platforms, research-grade network models.
Case studies
In production · Employer
·
Python / Microsoft Dataverse / Embeddings / FastAPI
ICP Engine & Data Enrichment
Context A German architectural glazing manufacturer sells through three distinct sales channels, each with a different ideal customer. Profiling was manual, static, and disconnected from the CRM.
System The engine reads customer records from Dataverse, verifies each company's online identity, and enriches records with digital-footprint and advertising-behavior signals. From the enriched base, channel-specific ideal-customer profiles are compiled on demand using analyst-selected filters — a living profile instead of a quarterly slide.
Architecture Python service against the Dataverse API; embedding-based similarity for profile matching — an approach first proven in a prototype that identified architecture firms likely to specify a specific timber product line; enrichment adapters feed a common schema. Output profiles are consumed directly by the lead-generation crawler.
In production · Employer
·
Python / Agentic workflows / Dataverse
Lead Generation Crawler
Context A profile is only worth what it finds. The channel ICPs needed to become pipeline without a research team doing manual prospecting.
System Takes a filtered profile from the ICP engine, scans the open web for matching companies, scores candidates against the profile, and writes qualified leads into the CRM with the evidence that qualified them attached. CRM feedback on lead quality flows back to sharpen scoring.
Architecture Agentic scraping workflow with per-source adapters, scoring against profile embeddings, idempotent CRM writes.
In production · Client work
·
Django / PostgreSQL / Railway
Logistics Operations Platform
Context A German logistics SMB ran its order lifecycle on spreadsheets and email: no roles, no audit trail, documents assembled by hand.
System Complete order management — creation through fulfillment — with role-based permissions for staff, customers, and management, and automated PDF document generation for the paperwork logistics runs on.
Architecture Django + PostgreSQL, deployed on Railway. Built solo, end to end: data model, permission system, document pipeline, deployment.
In production · B.Sc. thesis
·
Python / R / qgraph / bootnet / glasso / Neo4j / Streamlit
Computational Belief Modeling
Context How is a belief network structured — and how does that structure shift across contexts like institutional trust? The thesis modeled belief–affect–behavior structure in COSMO COVID-19 survey data.
System A reproducible Python/R pipeline estimates Gaussian Graphical Models (EBICglasso) across six survey contexts, computes bootstrap edge stability and focal-belief embeddedness, and encodes every network instance into a Neo4j belief knowledge graph for structured cross-context comparison.
Architecture Analysis in Python and R (qgraph, bootnet, glasso); Cypher import pipeline for the knowledge graph; public Streamlit app for interactive exploration including a Conditional Mean Shift explorer.
Outcome Graded 1.0. Live demo public.
GitHub ↗
Live demo ↗
Full index
-
In production
Agentic Fleet
A coordinated set of AI agents automating business-development workflows end to end — the company's first internal AI infrastructure.
Employer Python / LLM orchestration / Dataverse
-
In production
AdWatch — Ad Intelligence & Identity Verification
Monitors competitor advertising and verifies company online identities; its behavioral signals feed the ICP engine's enrichment layer.
Employer Python / FastAPI / SQLite / Apify / Meta Ad Library
-
In production
Retail Inventory & Tracking Platform
Inventory management and tracking system built for a retailer, covering stock movements and reporting.
Client work
-
In production
PredOpt Internal Operations Suite
The internal tooling our consultancy runs on: agent-driven workflows across CRM, operations, and marketing. Built for our own use — two founders, agents as the leverage.
Own venture n8n / LangGraph / Langfuse / PostgreSQL / Claude API
-
In production
PredOpt Website
Consultancy site with hub-and-spoke content architecture and an operations-diagnostic tool as lead generation.
Own venture Astro / Cloudflare Pages
-
In production
Content Automation Pipeline
Blog and newsletter pipeline driven by Telegram prompts: research, drafting, scheduling, and publishing run as one automated flow.
Own venture Postiz / Railway / crawl4ai / Claude API / Telegram
-
Prototype
Hybrid Vehicle Classification with Synthetic Data
Augmented a low-data vehicle classifier with Stable Diffusion synthetic images; the hybrid model reached 88.3% accuracy, beating both real-only and synthetic-only baselines. With Wilfred Okajevo.
GitHub ↗
Research PyTorch / ResNet-18 / Stable Diffusion
-
In production
Web Search Engine
Domain-restricted crawler with TF-IDF ranking, stemming, and fuzzy matching.
GitHub ↗
University Python / Flask / Whoosh
-
Prototype
Distributed Forum Platform
Hub-and-channel forum architecture with dynamically registered channel servers. With Ewa Godlewska & Franka Böckmann.
GitHub ↗
University Flask / SQLAlchemy / React
-
Prototype
CNN Interpretability with LIME
Model-agnostic feature-attribution pipeline across training checkpoints. With Marlon Dammann.
GitHub ↗
Research JAX / Flax / LIME / SLIC
-
Prototype
Feedforward Networks on CIFAR-10
FFNN implementation with per-class evaluation and weight-distribution analysis. With Zuzanna Bojarska.
GitHub ↗
University PyTorch
-
Prototype
Reversal Bandit Modeling
Rescorla-Wagner vs. Bayesian belief updating fit to human behavior; model comparison via BIC.
GitHub ↗
Research Python / NumPy / SciPy
-
Archived
Morefficiency Website
Company site for the LLC's consulting phase.
Own venture