My work.

Escaping vendor lock-in

When a critical vendor quadrupled its price, a technology company faced an existential risk. The challenge was to find a viable open-source alternative without derailing revenue-generating projects.

I led a structured evaluation, mapping five potential engines and directing a prototyping phase on the two finalists. To validate the winner, I led a customer-facing pilot project that created the blueprint for a new core architecture.

The outcome:

• Vendor risk eliminated: The company was no longer dependent on a single, unpredictable vendor.
• Immediate cost savings: A massive increase in licensing fees was avoided, freeing up capital for development.
• Production ready in months: The first client project on the new platform was live in production shortly after.

AI-powered content moderation

A global toy company's keyword blacklist was failing to stop cyberbullying on its digital platforms. I benchmarked several emerging NLP technologies, from which a Transformer model (M-BERT) showed the most promise for understanding nuanced, multilingual text.

The model performed well in the lab but failed on live data. This led to a critical insight: the problem wasn't the approach, but a lack of real-world training data. I solved this by building a data pipeline and feature store, retraining the model on a continuous stream of examples from the company's human moderation team.

The outcome:

• Increased moderator efficiency: The model pre-filtered a significant volume of inappropriate content, reducing manual workload.
• Improved platform safety: The system could now catch subtle forms of bullying that keyword filters miss.
• Introduced new capabilities: I successfully brought state-of-the-art NLP practices into the company.

Testing a new hardware frontier

When Apple announced the Vision Pro, a key question emerged: could this be the company's next hardware platform? As one of the first in Denmark with a developer unit, I was tasked with separating hype from reality.

A direct port was unviable, so I built a native proof-of-concept in one month using Apple's new APIs. This working demo became the primary tool for evaluation. By stress-testing the development pipeline with their 3D artists, I uncovered critical limitations in the tooling that made it unsuitable for their needs.

The outcome:

• Avoided a costly wrong turn: Evidence from the prototype prevented a strategic mistake and saved significant engineering effort.
• A clear, data-backed "no-go": The decision was based on a working model, providing clarity to recommit to the core platform.
• A high-value marketing asset: The prototype was so effective it became the company's primary demo for showcasing the potential of VR.

From a question to a product

My own project, SynthVR, started with a question: could VR make the hands-on experience of modular synthesis accessible? Success meant solving critical technical challenges - intuitive interaction and low-latency audio on standalone hardware - and the business challenge of validating an untested market.

I de-risked the technology by prototyping a novel interaction model alongside a custom audio engine. To validate the market, I launched a public alpha, and the community that grew around it guided the project from a tech demo into a refined commercial product.

The outcome:

• Validated an untested market: The project proved that a commercial opportunity for pro audio tools in VR existed.
• From R&D to ROI: A personal exploration successfully turned into a profitable commercial product.
• A proven technical blueprint: Created a novel high-performance audio engine that became the foundation of the product.

A high-volume hiring problem

A major Danish retailer needed to filter a high volume of applicants from a new, gamified survey. My task was to research and validate a machine learning model that could reliably predict a candidate's potential.

I ran a series of experiments comparing multiple approaches. The research showed that while complex deep learning models struggled, a simpler Random Forest solution was both effective and robust. I validated this finding on increasingly larger datasets before deploying the model as a production-ready API.

The outcome:

• A practical approach: I delivered a simple, effective solution that outperformed more complex alternatives.
• From research to production: I took the project from initial hypothesis to a deployed API integrated with the client's backend.
• Mitigated business risk: I provided the executive team with clear advice for managing algorithmic bias as the system scaled.