I build production AI systems that companies pay to deploy — computer vision pipelines, multi-agent LLM workflows, and document intelligence. Shipped across three continents.
I led the base map module — the pipeline that turns raw GPS coordinates into a clean, scaled, machine-readable road map — and independently built the agent system that pulls state-specific traffic regulations. The fine-tuned detection model and the downstream planning engine were owned by other engineers on the team; I worked closely with both, including helping prepare annotation guidance for the model training.
A US traffic engineering company produced temporary traffic control plans manually — an engineer studied a road, traced its geometry into AutoCAD, looked up the relevant state highway regulations, and worked out how many cones, signs and staff a job needed. Each plan took hours of skilled work and the process didn't scale with their order volume.
The goal: take a location, automatically reconstruct the road, and generate a regulation-compliant draft plan an engineer only has to review — not build from scratch.
Simplified intentionally — this shows the system's logic, not the client's internal infrastructure.
Real-world scale accuracy. A satellite snapshot has no inherent sense of metres. I pulled tiles at a fixed zoom and known ground-sampling distance, then stitched them so the final ~7K image had a consistent, calculable real-world scale — without that, every downstream measurement (lane widths, curve radii) would be wrong.
Resolution vs coverage. One API call couldn't give both a wide enough road span and the detail the model needed. I built tiling logic that captured overlapping high-zoom sections and merged them into a single seamless high-resolution canvas.
Vector fidelity. Model output (coloured masks for lanes, dashes, turn arrows, bike lanes, dividers) had to become clean AutoCAD geometry. I handled the conversion into DXF/DWG so curves and lane markings landed as proper engineering vectors, not pixel blobs.
50 states, 50 rulebooks. Traffic control rules differ by state and road type. Rather than hard-code them, I built an agent grounded to official state sources that returns a fixed-format answer set (work hours, permits, emergency contacts, staffing) for any given address.
Grounding beats hard-coding for anything that changes over time. The maintenance saving compounds.
For CV, input resolution often matters more than model choice. Fixing the image pipeline beat tuning the model.
Constraining an LLM to fixed sources + fixed output format is what makes agentic output trustworthy in production.
Design the human checkpoint deliberately. Knowing where not to automate is part of the engineering.
Open to remote AI engineering roles with EU and US companies. Comfortable with CET / EST timezones. Available for full-time positions and contract work.