|
I build AI systems end to end - the model, the pipeline that feeds it, and the product it ships inside. I like the whole stack of the problem: framing it, getting the data honest, training something that holds up out of sample, then wrapping it in an interface people will actually use. That has meant LSTM forecasters trading against live Binance data, Gemini-backed assistants with real conversation memory, and real-time collaborative systems on WebSockets. The other half of my time goes to open-source infrastructure. I am an experienced open-source contributor with sustained, merged work across CNCF security tooling, a Rust JavaScript engine, GPU virtualisation middleware and an Ethereum client.
|
ship it. break it. ship it again. |
| Project | What it is | Lang | My focus there |
|---|---|---|---|
| Kubescape | CNCF Kubernetes security posture scanner |  |
scan pipeline · SBOM · anonymiser |
| Boa | Embeddable JavaScript engine in Rust |  |
VM semantics · built-ins · Fetch runtime |
| HAMi | CNCF heterogeneous GPU virtualisation | |
vGPU metrics hardening |
| Besu | Ethereum execution client |  |
JSON-RPC trace & debug APIs |
What that work actually looked like
- Kubescape - scan pipeline internals: control filtering, output formats, SBOM emission, an MCP server surface, and the anonymiser that scrubs identifying data before results leave the cluster.
- Boa - VM completion-record semantics for
Yield/Return, ECMAScript built-ins (Iterator.concat,atob/btoa), and the WHATWG Fetch runtime includingAbortControllerandRequest/Response.clone(). - HAMi - hardened the vGPU monitor's metrics scrape path against panics on malformed or uninitialised device state.
- Besu - fixed unhandled exceptions and off-by-one validation in the JSON-RPC trace and debug APIs; enriched the readiness endpoint with diagnostics.
AI / ML
Languages
Data & Product
