AI/ML systems engineering
I work across agentic systems, GPU computing, neuromorphic exploration, and reproducible engineering—where an experiment should be as legible as its outcome.
input: scripted pathinput: your pointer
Inspect the live runtime Explore the labs
The static diagram explains the neuromorphic demo. The live visualization is an optional enhancement.
Neuromorphic demo
Every 50 ms the site turns your pointer, or a scripted path when nobody is interacting, into one [x, y, pressure] packet. kinetic-signals turns it into 16 features, axon-encoder encodes them as spikes, neuromod steps its LIF neurons, and synaptic-wiring owns the synapses and delays each pulse travels along. The adapter crate composes them into one WebAssembly module; the canvas only draws what it returns.
Before WebAssembly and WebGL start, and whenever they cannot, the drawing above is the same network: it is rendered at build time from the topology the browser bridge requires of every live snapshot. Native CUDA and FPGA results are recorded separately and never run in this browser, and recorded protocol data has its own page.
UNAVAILABLE · Rust/WASM
Telemetry reads the live Rust/WASM runtime. It needs JavaScript, WebGL, and WebAssembly; without them the static diagram is the complete demo.
Rows: NeuronId, top to bottom. Columns: steps, newest on the right. Marks: neuromod output spikes from the spike-event buffer. Bottom lane: axon-encoder spikes encoded per step, as a share of its 16 channels.
| Edge | Target | Weight | Delay | Polarity |
|---|
| Edge | Source | Weight | Delay | Polarity |
|---|
This contract does not export encoder input features.
Every value here except the input source is read from the live Rust/WASM runtime, from the same snapshots that drive the network view. The input source is the site's own DOM input tracker, which records whether the step's packet came from your pointer or the scripted path. Contract 5 routes axon-encoder spikes through synaptic-wiring into neuromod. The raster shows neuromod's output spikes, which are not fed back into the mesh, so a synapse listed here is the route the topology defines, not proof that its target received current. The runtime does not export a firing threshold or neuron-model parameters, so none are shown.
Research infrastructure
An AI/HPC research node and self-hosted engineering environment designed for hands-on systems work, experiments, and durable local workflows.
Data systems
Deterministic infrastructure for building provenance-focused synthetic engineering data, with an emphasis on executable sources and inspectable generation paths.
Research exploration
An ongoing exploration of quantization and neuromorphic, brain-inspired systems through careful research and prototype-oriented engineering.
Current focus
Designing engineering workflows that preserve their inputs, decisions, and evidence—while continuing to explore the boundary between efficient compute and brain-inspired architectures.