At STRV, I operated as an end-to-end ML engineer inside product teams. I did not just focus on building models, but was also making sure they shipped, scaled and behaved in production.

Much of my work revolved around search, recommendation, and retrieval systems. I designed hybrid search architectures combining Elasticsearch with vector embeddings, built RAG pipelines backed by vector databases, and implemented reranking layers to improve retrieval quality beyond naive semantic search. These systems powered real consumer applications, where latency and reliability mattered as much as model accuracy.

Beyond retrieval, I worked on recommendation engines using online learning techniques and fine-tuned computer vision models to solve cold-start problems. The common thread was pragmatic ML: tight feedback loops, measurable KPIs, and infrastructure designed for iteration.

I deployed most systems on AWS (Lambda, ECS, ECR, SageMaker, Bedrock), containerized with Docker, and supported by lightweight but effective MLOps practices. When needed, I stepped into data engineering (designing ETL pipelines with dbt and Meltano) ensuring that model performance wasn’t limited by messy upstream data.

This role sharpened my ability to think in systems: retrieval quality, serving latency, monitoring, infrastructure cost and product impact.

Key responsibilities:

• Researching and implementing RAGs, chatbots, and rerankers for enhanced retrieval performance.

• Developing recommendation engines and computer vision for mobile apps, addressing cold-start issues.

• Fine-tuning models for content moderation and audio projects, integrating with frontends and backends.

• Managing data engineering, migrations, and MLOps on AWS for production-scale solutions.

At STIL, I built an AI system in a medically driven environment with strict budget constraints, meaning no cloud, no overengineered stack but just careful engineering.

The core project was a tremor classification system using smartphone IMU data. I worked with both state-of-the-art time series techniques (MiniRocket) and strong traditional baselines (XGBoost), ensuring that improvements were measurable and clinically meaningful and not just technically impressive.

What made this role unique was its end-to-end ownership. I designed the FastAPI backend, deployed a PostgreSQL database (with indexing strategies and PGVector where needed), containerized everything with Docker Compose, and built CI/CD pipelines that automated testing, deployment, and even model retraining.

Because we ran on our own infrastructure, every architectural decision had cost implications. That forced discipline: efficient queries, lean services, clean abstractions.

It was a role that blended ML research, backend engineering, and DevOps in a context where reliability mattered as much as innovation.

Key responsibilities:

• Architectural ownership and technical decision-making.

• Production readiness and ML Ops considerations.

• Working with ambiguity and complex constraints.

• Enabling teams through knowledge sharing.

During my time at O2, I developed a customer life-cycle management model using TensorFlow neural networks to categorize customers, and a churn prediction system with XGBoost that achieved 4x better lift than random selection, enabling targeted call center interventions. As part of the PPF group, I extended these models to O2 Slovakia and Telenor in Serbia, Hungary, and Bulgaria. Data preprocessing pipelines were built in PySpark on our Spark cluster, later migrated to Azure, with automation via Jenkins and Azure Pipelines. Development leveraged Azure Databricks and ML Studio for efficient model building and deployment.

Key Responsibilities:

• Implementing neural networks and XGBoost for customer modeling and churn prediction.

• Expanding AI solutions across international teams in the PPF group.

• Building and migrating data pipelines in PySpark for large-scale processing.

• Automating workflows with CI/CD tools on on-prem and cloud infrastructure.