01 / Work
Built for
the real
world.
Six systems shaped by actual constraints: private data, enterprise scale, accessibility, developer friction, and the occasional unreasonable idea.
Selected systems / 06
- RateLab / limiter corePOST /api/checkFIXEDSLIDINGTOKENLEAKYREQUEST STREAM12DECISION429retry-after: 2s
01 / Distributed systems lab
4 algorithmsRate limits, under load
RateLab injects steady, bursty, and random traffic into deterministic limiters, then exposes every allow, deny, quota, retry clock, latency, and HTTP 429 decision in one inspectable interface.
- React
- TypeScript
- Vite
- Fastify
copilot / instruction
Add revenue to the chart and sort descending.
chat → schema → visual02 / Developer infrastructure
Chat → PBIXPower BI, directed by chat
An MCP surface over Power BI Desktop's report internals — add fields, rename columns, and reshape visuals from Copilot without opening the PBIX by hand.
- Python
- MCP
- Power BI Desktop
- TMSL / PBIX
- point / describe / promptship change
03 / Developer tool
7 IDEsPoint. Prompt. Ship.
Point at any element in your React dev server, describe the change in plain language, and vibefix ships a prompt — with the exact file, line, and expanded Tailwind classes — straight into Copilot Chat's Agent mode. No browser extension. No copy-paste. Published to npm.
- TypeScript
- Vite
- React
- Node.js
AA
contrast report12.4:17.8:104 / Accessibility
+12% retentionCertyFast, made accessible
A full-stack accessibility pass on CertyFast — React on the surface, .NET / C# and SQLite underneath — that lifted user retention 12% along the way.
- React
- .NET / C#
- SQLite
- WCAG 2.1 AA
contributions / spatialized
05 / Creative engineering
WebGPUGit City
Three.js + React Three Fiber, WebGL2 → WebGPU. Each developer becomes a procedurally-generated tower — floors are months, windows are daily commits, district placement is activity tier. In architecture and prototype phase.
- Three.js
- React
- TypeScript
- Node.js
- DOCSEMBEDFAISSANSWER
local models / private corpus
06 / Open source
Local-firstRAG without the cloud
Point it at a markdown corpus, get a persistent vector store and a chat CLI. Same commands whether you're on OpenAI, Ollama, or a local HuggingFace model. Built for teams that want RAG without shipping their docs to a SaaS.
- Python
- sentence-transformers
- OpenAI
- Ollama