Parhelion Yocto
Reproducible Yocto Scarthgap image line for DanSTAR CM4 flight hardware — custom meta-parhelion layer, staged WIC artifacts, and OS-vs-app release boundaries. Bench-validated on final OS 0.1 image.
projects and notes
Selected engineering projects and implementation notes.
Reproducible Yocto Scarthgap image line for DanSTAR CM4 flight hardware — custom meta-parhelion layer, staged WIC artifacts, and OS-vs-app release boundaries. Bench-validated on final OS 0.1 image.
RAUA Systems cloud ingest platform — local Docker E2E with FastAPI, Timescale, per-device auth, and Hub-ready routing. Co-founder architecture; Scaleway deploy documented, not shipped.
A Threadripper-based Proxmox server providing shared compute, GPU-accelerated CAD, and team services for DanSTAR.
A Go CLI that runs concurrent Linux diagnostics and emits token-aware Markdown reports for LLM context. Useful when debugging your machine.
DTU HPC coursework: a 3D fish-school simulation with CPU and CUDA backends, GH200 benchmarks, Nsight profiling, and ParaView export.
v1 homelab telemetry backend: API gateway ingest, PostgreSQL + InfluxDB + Grafana, devices-only boundary. Local E2E — not production cloud. Predecessor to RAUA Platform.
v1 ESP32 + SIM7000G firmware for bench and homelab E2E: deep sleep, batched uplinks, SD store-and-forward. Edge half of ZosteraGrids. Uplink JSON is still RAUA Platform's ingest contract reference—not production field deployment.
A Python autonomous robot control stack on a Raspberry Pi. Hardware talks over MQTT.
DanSTAR flight-computer program: team-owned Python on Raspberry Pi and an in-house HAT. Explicit mission state, safety interlocks, and linked OS and operator workstreams.
Operator checklists and release engineering for DanSTAR flight software documentation.
Master's thesis: design and formal evaluation of a local, modular voice assistant with self-hosted language models, local speech services, and episodic memory.
Distributed self-hosted voice assistant — STT and TTS on edge hosts, GPU-backed inference and PostgreSQL+pgvector memory on a server, with a formal thesis benchmark campaign on x86 and ARM.
Spiking neural network components in PyTorch: stateful temporal modeling, surrogate gradients, and vibration classification.
A feed-forward neural network in NumPy only, with manual backpropagation, optimizers, and WandB experiment tracking.
A C++ search client and benchmarking setup for cooperative multi-agent planning, focused on deterministic execution and performance analysis.