Diya Hussein
Six systems, built close to the metal. One threshold, and no way back the way you came.
Enter the workIt was never locked.
Only closed, and waiting to be walked toward.
AI/ML & Systems Engineer
Work
Six projects · MIT
AI / ML
02
01
MeasuredYOLOv8n · MB
ONNX quantized · CPU only
Real-time edge AI pipeline — YOLOv8 + ONNX quantization, FastAPI, Grafana monitoring. 28 FPS on CPU.
02
StructuralFrom scratch
Multi-head · flash attention
GPT-style decoder built from nothing — multi-head attention, flash attention, WikiText-2 training.
Systems
02
03
Measuredµs added
p50 · 1-byte roundtrip
Async L4 TCP proxy with io_uring kernel bypass. Windows loopback; Linux ~2–3× faster.
04
Measuredops/s
Order book market order
Event-driven backtesting engine with limit-order-book simulation and Python bindings.
Trust & Verification
02
05
SchematicWhat did it gain?
Version diff · 10 rules
Trust registry for MCP servers. Catches the package that stays clean six versions, then turns.
06
StructuralEd25519 · X.509
Root CA → signed leaf
Minimal X.509 PKI — self-signed root CA, leaf issuance, Ed25519 signatures. Built to demonstrate.
Capabilities
Eight areas
AI / ML
Bilingual AR/EN/RO intent and slot-filling, strict-JSON single-call pipelines, offline and local inference.
Tool-use registries, terminal agents, multi-tenant deployment.
Real-time object detection, ONNX quantization, edge deployment, model benchmarking.
Transformers from scratch in JAX, attention, training.
Systems
io_uring kernel-bypass I/O, lock-free data structures.
L4 proxying and event-streaming paths.
Order-book simulation, backtesting, Sharpe/Sortino metrics.
X.509 certificates, Ed25519 signatures, trust chains.
About
Timișoara, Romania
I work across two layers of the stack, and I care most about the seam where they meet.
Up top: AI systems — agents, ML pipelines, bilingual LLM applications, and real-time inference. Down low: systems engineering — low-latency networking, order-book reconstruction, and the cryptographic primitives that keep things correct and trustworthy.
Most of what I build is infrastructure, which means it has no screenshots. What it has instead is numbers and a method you can rerun. Every figure on this page comes from a benchmark in the repo it belongs to — and where there is no measurement, the card says so rather than implying one.
Contact
Open to work
If any of this is the kind of thing you need built, I would like to hear about it.