My projects, roles, and technologies.

• Mapped the risks of uncontrolled agentic systems, including shadow AI, missing accountability, weak auditability, and violations of least privilege.
• Designed an identity lifecycle and distributed ownership model for agents, capabilities, and data, with governance intensity tied to risk.
• Defined a four-plane target architecture spanning IAM control, policy decisions, enforcement, and controls inside target systems.
• Specified and validated an Agent Identity MVP through Microsoft Entra Agent ID, plus an MCP server MVP grounded in a concrete use case.
• Delivered the requirements catalogue, recommendations, and roadmap for the implementation phase.

Tools used

• One Identity
• Entra Agent ID
• MCP
• dmGPT
• Microsoft Office

Tools used

• AWS Lambda
• API Gateway
• DynamoDB
• Terraform
• React
• Vertex AI

• Designed a fully serverless AWS backend for a consumer mobile application.
• Implemented the complete infrastructure as code and integrated OpenAI alongside services outside AWS.
• Prototyped LLM prompts and model behaviour, then built monitoring and alerting for production operation.

Tools used

• AWS CDK
• Lambda
• DynamoDB
• Python
• EventBridge
• SNS

• Founded and led a startup focused on making ML productionisation repeatable and faster.
• Compared and prototyped cloud-native deployment architectures before implementing a scalable serverless platform on AWS.
• Created a reusable repository template, a dedicated Python package, and automations for new projects.
• Built the full operating stack: CI/CD for code and models, staging environments, package and artefact management, monitoring, and alerting.
• Produced a practical online course for teams adopting q·ops and AWS-based MLOps.

Tools used

• AWS CDK
• AWS Batch
• Lambda
• Metaflow
• CloudWatch
• Python

• Structured user stories and business use cases across several operational domains.
• Defined platform capabilities and evaluation criteria for potential IoT solutions.
• Interviewed stakeholders and engineering teams to surface the limits of existing and planned IoT projects.
• Compared make, buy, and hybrid options across on-premise, vendor, and cloud platform approaches.
• Developed AWS and GCP architecture concepts, prioritised the work, and produced an implementation-ready roadmap.

Tools used

• AWS
• GCP
• IoT
• draw.io
• Miro
• Microsoft Office

• Derived platform capabilities from the existing AI portfolio and prior delivery experience.
• Assessed current platform and deployment solutions against a detailed maturity model.
• Identified the technical and organisational barriers preventing teams from productionising AI reliably.
• Designed staged AWS architectures for lightweight and full-scale model serving, updating, and retraining.
• Converted the assessment into tooling recommendations, priorities, and a practical development roadmap.

Tools used

• AWS
• draw.io
• Microsoft Office

• Joined high-frequency pressure-sensor time series with damage labels from SAP.
• Cleaned and validated measurement and label data, then engineered features for a Random Forest baseline.
• Reduced dimensional complexity by roughly 64% with an autoencoder.
• Developed a robust rolling RANSAC method for earlier wheel-damage prediction with stronger outlier resistance.
• Prepared the codebase and documentation for continued work by the project team.

Tools used

• Python
• PyTorch
• scikit-learn
• AWS

• Researched recurring MLOps effort and requirements with teams across the organisation.
• Designed a standard AI platform using open-source components and AWS-native services.
• Evaluated model-serving and automated retraining approaches, then iterated an AWS resource blueprint for future deployments.
• Validated permissions, service boundaries, and security decisions with the cloud engineering team.
• Delivered detailed technical documentation and stakeholder-ready architecture decisions.

Tools used

• SageMaker
• S3
• Lambda
• Airflow
• DVC
• Python

• Defined the functional requirements for named-entity extraction from documents.
• Designed a human-in-the-loop workflow to improve extraction quality iteratively.
• Built a Dash wrapper that integrated into a React frontend and connected the application to the backend through an API.

Tools used

• Python
• Dash
• FastAPI
• OpenAPI

• Led structured interviews and workshops with chemists, stakeholders, and prospective users.
• Translated scientific and user requirements into user stories and journeys.
• Reviewed market solutions and reference applications to shape the UX direction.
• Created the application mock-up with UX and UI designers, then defined the high-level technical architecture with DevOps specialists.
• Produced the request for proposal for the implementation phase.

Tools used

• Figma
• Microsoft Office

• Assessed the risk of reconstructing sensitive training data from drug-discovery models.
• Mapped membership inference, model extraction, model inversion, property inference, and molecular reconstruction attacks to the customer context.
• Ran selected high-relevance attacks using public bioactive molecule databases.
• Built a Python library for repeatable model-security assessment.
• Recommended and implemented technical safeguards in the customer software.

Tools used

• Python
• Pandas
• scikit-learn
• KVM

Tools used

• BERT
• ECS
• Lambda
• DynamoDB
• CodeBuild
• Python

• Designed the scalable AWS architecture for production email classification.
• Led the engineering team, owning technical planning, prioritisation, and delivery.
• Brought the preprocessing pipeline and Random Forest model to production quality.
• Implemented stateless containers for preprocessing and inference, then integrated the surrounding AWS services.
• Presented the work to the customer and later at the PyData and NLP Meetup Hamburg.

Tools used

• Random Forest
• ECS
• Lambda
• API Gateway
• S3
• Python

• Analysed and prioritised six NLP use cases with business teams and subject-matter experts.
• Integrated and harmonised heterogeneous datasets from six parts of the organisation.
• Built complete prototype loops from preprocessing and feature engineering through model training and evaluation.
• Selected model approaches against each business objective and target variable.
• Delivered management-ready results on a two-to-three-week cadence per prototype.

Tools used

• SageMaker
• Python
• Pandas
• scikit-learn
• Jupyter

• Prepared, converted, and analysed high-frequency infrastructure sensor data.
• Developed an unsupervised anomaly-detection proof of concept.
• Identified potentially critical events and created visualisations for engineering interpretation and decision support.

Tools used

• Python
• Pandas
• scikit-learn
• Jupyter

• Prepared and analysed unstructured text from email and Jira tickets.
• Developed topic-clustering methods to organise large text corpora.
• Classified and filtered email by business relevance.
• Used exploratory analysis to turn text patterns into business cases and decision support.

Tools used

• Python
• scikit-learn
• Jupyter