I build
Senior software engineer, 8+ years. I take systems from a blank repository to production — LLM agents that test real browsers, computer-vision pipelines that count people in live video, and cloud-native .NET platforms running inside hospitals and councils.
I like the hard middle of software — the part where a clever prototype has to become something a team can run on a Monday morning. Most of my work sits at the seam between AI and real systems: model output that has to be deterministic, agents that must not burn tokens on every run, vision pipelines that keep frame rate on ordinary hardware.
Recently that has meant three products built end to end — OpenAgentQA, an agentic browser-test runner that records locators once and replays them for free; AiUML Studio, a Blazor WebAssembly diagram editor with a custom SVG engine and no third-party diagramming library; and HTML to Video Studio, a deterministic text-to-MP4 renderer where identical input produces byte-identical output.
Before those: computer vision for crowd and traffic analytics, embedded AI on a smart helmet, and enterprise .NET platforms for healthcare catering and financial document processing.
.NET 8–10, C#, Blazor, React, Next.js — API to pixel.
LLM agents, RAG, tool-calling, MCP, local inference.
YOLOv8, OpenCV, tracking, zone logic, live streams.
Azure AI, Document Intelligence, PostgreSQL, MSSQL.
Depth where it counts, breadth where it helps. Bars are honest self-assessment, not decoration.
Nine production systems across AI tooling, computer vision, healthcare, finance, mobility and IoT. The first three are current flagship builds.
Write browser tests in plain English. An agent reads a screenshot and decides what to click; the first run records ordinary Playwright locators, and every run after replays them with no model in the loop.
A Visio-like editor in Blazor WebAssembly. The diagram engine is hand-built — plain SVG and one thin JS interop module, no third-party diagramming library. Animated connections survive export to a standalone SVG.
Describe a video; an AI writes an HTML composition with a real timeline; the studio renders it deterministically to MP4 with headless Chromium and ffmpeg. Identical input produces byte-identical output. Pure .NET, no Node.
Local-first research engine using tree-based vectorless RAG for deep analysis of complex PDFs. Hierarchical document mapping with multimodal indexing across text, images and charts.
YOLOv8 detection and tracking for surveillance and crowd control. Interactive polygon zone drawing, IN/OUT directional logic, annotated video output and CSV/Excel export.
Directional vehicle counting dashboard — polygon zones, real-time tracking and class-specific analytics for cars, buses and trucks, with exportable traffic statistics.
Firmware and AI modules for a helmet streaming live video with onboard person and object detection, plus a real-time stream server for remote monitoring.
React/Next.js front end, C# APIs, SQL schema work and Azure Storage integrations for a catering management platform deployed across hospitals, councils and public institutions.
C# and PostgreSQL backend for financial document processing, with Azure AI Document Intelligence extracting receipts and invoices into structured CPA reports.
Hover to pause the rows.