Work
Systems I've designed and built
Each of these started as a real problem — a workflow held together by spreadsheets, an inbox with no structure, or an infrastructure question with no clean answer. Every project has a full case study behind it.
Flagship Work
The three worth reading first
End-to-end systems where I owned the problem, the architecture and the implementation.
Replacing spreadsheet-driven inventory with a centralized operations platform
FAB Construction Inventory Management System
- The problem
- A construction company tracked stock, tools and job status across spreadsheets and informal messages — so nobody could reliably answer what was in stock, where it was, or which work was blocked.
- What changed
- One system for products, multi-location stock, suppliers, stock movements, work processes and role-based access, replacing the spreadsheet-and-messages workflow across FAB's sites.
Automating unstructured IT support intake and ticket routing with AI
AI-Powered IT Support Agent
- The problem
- IT requests reached RYCO's helpdesk as ad-hoc emails and Teams messages with no categorization, routing or SLA tracking, so every request began with manual triage.
- What changed
- A guided conversation that turns a vague message into a validated, categorized, routed ticket with an automatic confirmation — deployed in Teams at RYCO, and rebuilt as a public live demo carrying no organizational data.
Choosing a parallelization strategy on measured runtime and energy, not intuition
High-Performance 1D Earth Mover's Distance
- The problem
- Distribution-comparison workloads do not scale sequentially, and the strategy that looks fastest on paper is often neither the fastest nor the cheapest one in practice.
- What changed
- Measured on Basel's SciCORE cluster: 73x for 1D EMD's compute phase on a single A100 and 4,245x for matrix multiplication on MPI+CUDA — alongside one case where OpenMP used ~91x less energy than the distributed version.
Applied Research
Deeper technical work
Research-driven work where the deliverable is a measured answer rather than a shipped platform.
Measuring what formal privacy guarantees actually cost a credit-risk model
Privacy-Preserving Credit Default Prediction
How much accuracy a credit-risk model loses under real privacy guarantees — measured with membership-inference attacks rather than assumed.
Contact
Let's talk.
Hiring for AI, data or platform engineering — or building something technically difficult? Tell me what you are working on.
- AI & data engineering
- Forward-deployed / customer-facing engineering
- Distributed systems
- Applied AI & automation
- Data platforms