I void
warranties.

This is a project to ingest public safety communications within the Washington DC area, transcribe them in real-time, and use an AI agent to categorize and manage a real-time incident board for interested parties to keep updated on public safety concerns as they happen.

A Whisper model performs the transcription, with an additional workflow for evaluating different transcription models or settings. I additionally use entity extraction combined with DC street names and other vocabulary to improve signal, provide contextual links, and allow for geofenced notifications. The incident board management agent has access to tools allowing it to research transmissions and use audio-capable reasoning model to get a second opinion on transcripts or verify its understanding.

Using NATS JetStream and an optimized SDR -> WebSockets path I’m able to measurably play transcoded audio samples in the browser less than a second after RF capture, with transcripts usually following within a few seconds (and often competing).

The system also provides a performant and expressive search mechanism and an archive of transmissions going back roughly a year.

The custom 20-inch Maksutov at Vega-Bray had been dark for ~15 years, its original control electronics gone and no documentation. We reverse engineered the hardware, and built an ESP32-based LX200-compatible controller. On top of that, a Go-based API serves a React frontend providing a celestial map and user interface for the telescope. We leverage a 3D model and ROS-inspired URDF model and kinematics to build motion plans that avoid self-collisions, a real concern with these German equatorial mount telescopes.

I additionally built a simulated telescope so that the whole system could be tested end-to-end, including firmware, before deploying to real hardware. This drives the demo here.

This is an ongoing project to build an orchestration layer for Software Defined Radio. I’m using Kubernetes Dynamic Resource Allocation (DRA) to manage hardware SDR resources, configuring the radios based on CRD resource requests, and streaming data through shared memory segments to unprivileged consumer pods.

A proof of concept exists working end-to-end. Workloads will be able to ask for abstract SDR streams “20 MHz around 2.4GHz”, and the platform will satisfy that request by re-using an existing compatible stream, allocating an idle SDR, channelizing a shared-use SDR, or aggregating two or more SDRs within a coherency domain.

I’ve achieved single-copy transfers with shared memory segments and loans, with a path to zero-copy DMA-to-processing options in the future.

Luna is a 1-meter lunar relief sculpture of Tycho crater.

The starting point was gigabytes of NASA elevation data — SLDEM-2015 lunar topography, delivered as high-bit-depth GeoTIFFs. I wrote a Python pipeline to ingest it, sample and rescale the heightfield, and emit tiled OBJ meshes sized to a printer bed. The pipeline handles smoothing and depth-scaling so the relief reads at viewing distance without exaggerating the Moon’s already-subtle topography.

From there: printed in PETG on a hardboard backing, assembled, mounted.

Projection mapping completes the effect, letting me control shadows and stylization.