10/07/2026 updated

**** ******** ****
100 % available

Machine Learning Scientist & Software Engineer | AI Systems | LLM | Deep Learning

Malmö, Sweden
Worldwide
Abitur
Malmö, Sweden
Worldwide
Abitur

Profile attachments

CV-Nouri.pdf

Artificial IntelligenceMachine LearningTensorFlowPyTorchLarge Language ModelsDeep LearningScikit Learn
Maschinelles Lernen und Software seit über 20 Jahren; Schwerpunkt heute: LLM- und Agentensysteme in Produktion; von der Bewertung (LLM-as-a-Judge, Holdout-Verfahren) bis zum sicheren Rollout. Projektarbeit läuft über meine schwedische Aldeology Consulting AB, Remote aus Malmö (CET).

LLM- und Agentensysteme in Produktion
  • NL-to-SQL-Agent (GPT-4o) für rollenbasierte, natürlichsprachige Abfragen auf Finanzdatenbanken, mit LLM-as-a-Judge-Bewertung (Securitas)
  • Selbstverbessernder LLM-Verifier für Nachhaltigkeitsdokumente: 98,5 % Genauigkeit gegen 84,6 % Human-Baseline (66 Dokumente, 11.800 Bewertungen); Experten-Feedback fließt direkt in Prompts und Kontext zurück (BIMobject)
  • Agentenplattform für quantitative Forschung mit acht spezialisierten Agenten: Orchestrierung per State Machine, Kampagnen-Gedächtnis, ehrliche Stop-Regeln, Keep-or-Discard über frozen Holdouts; 188 Experimente in 14 Kampagnen
  • OSINT-Aggregator: LLM-Clustering, Zusammenfassung und Relevanz-Scoring über tausende Quellen, Deployment auf Azure (Securitas)
Coding-Agenten und Entwickler-Tooling
  • Autor von Pilish, einem Emacs-Frontend für den pi Coding-Agenten (auf MELPA, 10 externe Beitragende); Skills und Erweiterungen für Claude Code und Codex CLI
  • claucker: Docker-Isolations-Wrapper für den autonomen Betrieb von Claude Code (Dateisystem-Isolation, entzogene Capabilities; bewusst ohne Netzwerk-Isolierung; kein Sicherheits-Sandbox)
  • Selbst gehostete Modelle: Ollama als Integrations-Test-Backend für Pilish (Docker, CI); Test-Infrastruktur, kein Produktiv-Serving
Klassisches ML, Bild und Edge
  • Medizinische Bildgebung: vier erteilte US-Patente für Ultraschall-Guidance; Echtzeit-Landmarken-Tracking mit 2,83 mm mittlerem Fehler (MICCAI 2015)
  • Edge/IoT: C++-Framework auf Apache-TVM-Basis; H.264-Motion-Vector-Analyse auf ressourcenbeschränkten Geräten (SmartThings/Samsung)
  • Kaggle Master; frühe CNN-Anwendungen auf Audiosignale (Walrufe, unter 245 Teilnehmern Top 3)
Engineering
  • Python seit 2001 produktiv; C++, SQL; PyTorch, TensorFlow, scikit-learn, DuckDB, Spark
  • AWS, Azure, GCP; Docker, Linux, CI/CD; automatisierte Tests und Code-Review-Kultur
  • Architektur und Führung: AI-Team-Aufbau bei 4Catalyzer (3 auf 6 Personen) und Lotik/Samsung NEXT; LLM-Workshops und Mentoring
Open Source und Community
  • skorch: scikit-learn-kompatibler PyTorch-Wrapper, in über 1.500 Repositories verwendet; nolearn (951 Sterne); Lasagne-Core-Contributor
  • Gründer des Berlin Machine Learning Meetup (über 9.000 Mitglieder)
Sprachen: Deutsch (Muttersprache), Englisch (fließend)

Experience

Professional experience10+ years
Freelancer since2001

Languages

GermanNative speakerEnglishFluent

Project history

Co-founder

Stealth legal-tech startup
Building a matter-centered AI workspace for German law firms with page-level citations, transcript timestamps, and drafting in a Tiptap-based editor. Designing for GDPR and professional secrecy with immutable originals, server-controlled matter scope, and deletion that survives backup restores.

AI Consultant

Aldeology Consulting AB
Built an LLM-agent platform running quantitative research unattended with eight specialist agents under an orchestrating state machine, campaign memory, honest-stop governance, and a frozen-holdout ladder; 188 experiments across 14 campaigns. Rebuilt data core clean-room on 187 GiB of market-by-order tick data with order-book reconstruction proven against vendor snapshots.

Senior AI Engineer (Contractor)

BIMobject
Built a self-improving LLM verifier for sustainability documents achieving 98.5% accuracy vs 84.6% human-validator baseline across 66 expert-annotated documents and 11.8k judgments. Improved an 839-prompt human-review compliance checklist with expert annotations, cutting AI errors by 45%. Created a feedback loop where expert-reviewed failures automatically guide updates to prompts, context, and applicability logic.

Lead ML Engineer and Coach (Contractor)

Securitas Group
Led development of an agent AI system (GPT-4o) for secure, role-based natural-language querying of financial databases (NL-to-SQL) with LLM-as-a-Judge evaluation. Architected an AI news aggregator processing thousands of websites for OSINT with LLM topic clustering, summarization, and relevance scoring on Azure. Drove LLM adoption and engineering rigor across the AI department through workshops, mentoring, code-review, and CI/CD discipline.

Lead Principal, Systems Engineering

AT&T
Built WebRTC audio-stream forwarding for speech and emotion recognition in a new call-center product; evaluated third-party speech systems and designed the supporting ML cloud architecture.

Machine Learning Scientist (Contractor)

SmartThings
Built a detector combining H.264 motion vectors with neural networks for on-device inference on resource-constrained IoT devices. Created core C++ framework based on Apache TVM for edge ML model delivery supporting high efficiency, model chaining, and multi-platform deployment. Provided project leadership and mentoring to the data science team.

Consultant, Machine Learning Scientist, Trainer

Natural Vision UG
Directed a boutique machine learning consultancy managing up to 5 people, leading client projects from discovery through productization and operations. Delivered solutions using deep learning and traditional ML (CV, NLP, Predictive Analytics), including building a foundational ML framework for Otto Group and evaluating XING's news feed algorithm. Lectured professionally at bootcamps via Data Science Retreat.

Chief Scientist

Lotik Labs at Samsung NEXT
Built and led the startup's Data Science team; architected a novel ML solution combining signal processing with deep learning for embedded leak detection on low-power devices.

Head of AI (prev. Contractor)

4Catalyzer
Progressed from contractor to leading a multi-disciplinary AI team building medical imaging (ultrasound) applications. Invented a patented ultrasound guidance system using heavily customized U-Net architecture, enabling non-experts to capture diagnostic-quality images. Collaborated with doctors and product teams, building custom models and data acquisition/annotation systems.

Software Engineer, Consultant, Trainer

Self-Employed
Built a quality management system (Python/Pyramid) used by 1,900+ local authorities across Europe serving 77M+ people. Led technical development for vienna.info (Plone/Varnish), one of Austria's highest-traffic websites.

Contact form

Log in to get in touch

You need to be logged in to use the contact form.

Sign upLog in