Engineering

AI enablement, strategy and cross-org delivery

• Core contributor to the AI Center for Enablement, acting as an internal AI consultant — defining how AI is scoped, delivered and operationalised across HR, L&D, Operations, Finance, Global Supply Chain, IT, Sales Operations, R&D and Logistics.
• Led AI initiative intake and use-case discovery: problem definitions, feasibility assessments, roadmaps and execution plans across multiple domains.
• Bridged business leadership, product owners and platform teams to align AI initiatives with strategic objectives and data readiness.

Platform and MLOps ownership

• Drove adoption of scalable MLOps patterns across Azure ML and Databricks — MLflow-based experimentation, model registry practice, Vector Search, RAG architectures and Unity Catalog.
• Authored internal standards and guidelines for CI/CD for ML, model lifecycle management and deployment readiness.
• Worked with DevX and platform teams to improve reliability, monitoring, reproducibility and governance of production ML systems.

Agentic AI, conversational systems and automation

• Led design and delivery of several production agent-based systems on Microsoft Copilot Studio, Databricks and Azure, serving customer-facing, sales, operational and HR audiences. Representative examples:
• Customer-facing product discovery — a public web assistant grounded in product databases, technical documentation and marketing content, supporting product selection and customer support.
• Manufacturing diagnostics — a shop-floor assistant for diagnosing machine and process errors, built with domain experts and grounded in troubleshooting documentation.
• Sales enablement — an assistant surfacing verified facts, references and case material to support outreach to prospective customers.
• Learning and development — an HR assistant mapping competency gaps and recommending personalised learning paths against an established 70/20/10 framework.
• Integrated agents with Power Automate, evaluated trade-offs between traditional RPA and agentic automation, and documented licensing, governance and operational best practice.

Product intelligence, generative AI and content automation

• Led an AI-driven product description initiative generating short and long descriptions across a catalogue of 35,000+ products using LLM-based pipelines.
• Designed and implemented scalable data and generation workflows in Databricks and Azure, integrating structured product data with unstructured marketing content.
• Defined extensive business rules, quality guidelines and validation logic to ensure brand consistency, regulatory compliance and output quality.
• Collaborated with product, marketing and IT to operationalise live content generation and continuous updates.

Industrial computer vision — automated tool wear inspection

• Technical lead, owning the end-to-end computer vision pipeline from data annotation and dataset strategy through model development to cloud deployment.
• Designed and deployed object detection models (YOLO, Faster R-CNN) for automated visual inspection, replacing a manual inspection process.
• Drove large-scale annotation efforts: labelling guidelines, annotation workshops, quality assurance and dataset versioning practice.
• Contributed to the architecture blueprint, risk management, and functional and non-functional requirements, including migration of AI workflows to cloud infrastructure.

Reliable adaptive predictive maintenance and intelligent decision support With SCANIA CV AB and Linköping University

• Collaborated with cross-disciplinary teams — applied scientists, software engineers and analysts — to deliver ML solutions tied to business KPIs.
• Led development of a library and proposed MMK-means, a real-time inference clustering method handling up to 50% missing data while matching full-data K-means accuracy.
• Designed a novel cost-sensitive Transformer model, cutting truck braking system costs by 15%.
• Oversaw implementation with a team of engineers and research scientists.

Mitigating toxicity and bias in large language models via reinforcement learning With SINTEF AS and KTH

• Developed a model-agnostic RL framework minimising toxicity across LLaMA, GPT, BERT and RoBERTa architectures.
• Evaluated with diverse toxicity metrics, achieving significant reductions in toxicity, profanity, sexually explicit content, flirtation, threats and insults.
• Ran fairness and bias mitigation experiments, demonstrating improved inclusivity and fairness scores in generated output.

AI-based control and coordination for smart converters With Hitachi Energy and KTH

• Developed a multi-agent deep RL (MADDPG) system for smart converters.
• Enhanced actor-critic models with Accelerated-TD3, achieving a 200% improvement in learning efficiency by integrating prior knowledge.
• Pioneered a communication-efficient federated learning framework for MARL, improving collaboration efficiency 40×.
• Investigated finite-sample convergence of federated algorithms including error-feedback FedProx and FedAvg, achieving 80% faster convergence without data similarity.
• Examined non-asymptotic convergence of distributed momentum methods under biased gradient estimation from compression and clipping.

Deep reinforcement learning for sustainable power systems With KTH and the University of California

• Created a PPO-based deep RL model for sustainable power systems, reducing energy consumption by 16%.
• Designed a new Gymnasium-based environment for evaluating multi-agent RL algorithms, shortening development cycles.
• Introduced TA-Explore, a novel exploration strategy accelerating convergence in simulation by 150%.

Teaching and supervision

• Supervised 30+ Master's students in AI and Health Informatics at Stockholm University and Karolinska Institutet, applying deep learning, Transformers, multimodal and reinforcement learning across healthcare, forecasting, recommender systems, network optimisation, planning, agriculture and fairness.
• Developed a new MSc course in Reinforcement Learning — built the teaching material from scratch and designed and ran five laboratory sessions for the first cohort. 187 students.
• Course instructor — Current Research and Trends in Health Informatics, focusing on time-series analysis, deep learning methodology and reinforcement learning. 91 students.

Fraud detection and risk assessment

• Led an ML team of six for over six months, working with product, design, data scientists and engineering managers to scope initiatives using ML — including problem definition specifications and stakeholder management.
• Built pipelines for collection, preprocessing, feature engineering, training, evaluation, deployment and batch inference, cutting false positives 40%, avoiding roughly $1M in losses and reducing implementation time 30%.
• Created a zero-to-one specification for comprehensive risk assessment across multiple business entities.
• Built prototypes and demonstrations using MLOps principles — CI/CD pipelines, automated workflows and containerisation with Kubernetes and Docker — to streamline deployment, monitoring and troubleshooting of production ML at scale.

Customer churn prediction

• Led initiatives to identify performance gaps, then co-designed and rolled out a system using CatBoost, LightGBM, XGBoost and Random Forest, improving prediction accuracy 12%.
• Addressed data imbalance with SMOTE and ADASYN, producing more robust performance on minority classes.
• Contributed content, teaching and tutoring for internal data and ML training programmes, and led ML reading groups across teams and product areas.

Supply chain optimisation

• Built and serviced an order lead-time prediction system using linear regression, decision tree regression and gradient boosting, reducing lead-time variance 25%.
• Scaled ML deployments for large-scale data processing and transactions, troubleshooting deployment issues to hold 99.999% uptime.
• Onboarded team members from junior to staff level and mentored engineers on leading initiatives — problem scoping, test specifications and defining deliverables.