Projects
Operating systems, inference runtimes, simulators and tools, most of them born from a research question or a problem at work. Star counts are live from GitHub.
AI-native operating system for Cortex-M33 microcontrollers with on-die NPUs, built on Zephyr for the NXP FRDM-MCXN947. Inference is the primary workload: tensor-aware memory arena, NPU and DSP abstraction, model store with OTA, dual-core inference offload.
Zero-allocation C11 library for neural network inference on ARM Cortex-M and Cortex-A. Eighteen operators across Dense, Conv, LSTM, GRU and Transformer layers, INT8 with INT32 accumulation, a deterministic graph runtime and a sub-50 KB flash target.
Distributed, adaptive resource orchestrator for AI workloads on clusters of embedded Linux nodes. Real-time resource monitoring from procfs and sysfs, peer discovery over UDP, three scheduling modes, thread and memory pools.
Memory-aware runtime for partitioned Transformer inference on devices with a few gigabytes of RAM. Type-state inference engine, RAII buffer guards, sequential, greedy and speculative partition planners, resource monitor.
Header-only C++23 state-estimation framework built for vessel navigation. EKF, UKF and complementary filters behind a single concept, Mahalanobis outlier gating, compile-time sensor composition, 268 tests under strict warnings.
The experimental platform of my thesis. Five Raspberry Pi 4 boards on an isolated 802.11ac network and an edge server, driven end to end by the NNPM web application: scenarios, concurrent SSH dispatch, per-layer profiling under tc-netem shaping and compute load, INA260 power traces, the RR and SJF delay model and partition optimisation. Runs without hardware in local mode.
Reference implementation of the SPAWC 2026 paper. Each layer's attention heads are scheduled across devices so attention runs in parallel everywhere, and a speculative intermediate output is forwarded before the FFN completes. PyTorch RPC over WiFi on Raspberry Pi clusters.
Discrete-event simulator for resource-aware decoder-only Transformer inference across heterogeneous edge clusters. Models devices, attention heads, projections and FFNs, growing KV caches, and compares EdgeShard, Galaxy and the resource-aware algorithm from the WiOpt 2025 paper.
Raspberry Pi testbed for layer-level DNN partitioning across a WiFi network. Greedy, simulated annealing and genetic optimisers, per-layer latency and power profiling with INA260 sensors, latency injection and a Flask dashboard. Demonstrated at ACM MobiHoc 2024.
Linear, LSTM, GRU and Transformer forecasters that predict video quality of experience from network metrics, with dataset generators, hyperparameter tuning and an inference middleware. Built for the TARGET-X tele-operation work.
Ryu controller application that mirrors traffic to Snort, blacklists source and destination pairs on alerts, and installs high-priority drop rules on the OpenFlow switch.
Simulators for reactive firewalling in software-defined mobile ad hoc networks, from the SMOTANET cybersecurity work package.
Systematic comparison of threshold, linear, DNN, LSTM, GRU and Transformer detectors for transport-layer DDoS on IoT and 5G gateways, measured on accuracy, latency, memory footprint and early-warning lead time.
Run the real stack, prove it before the field. Ground, aerial and surface robots run their unmodified ROS 2 (and bridged ROS 1) codebases in isolated containers against Gazebo, orchestrated by a scenario engine and observed from a browser-based 3D cockpit.
Free NMEA 0183 and Signal K data-stream simulator for Windows, macOS and Linux. Simulates a vessel under way with GNSS, heading, depth, wind, autopilot, AIS and propulsion sentences over serial, TCP, UDP and WebSocket, with a headless mode for CI.
Back up an Ubuntu computer to an external drive and rebuild it on new hardware or a new release: files, apps, settings and system configuration, with LUKS and VeraCrypt support and a GTK4 interface.
Screening-level decision support for nature-based urban cooling, developed for Criterra. Scores heat priority, cooling potential, cost and equity for 121 NbS typologies, with map-based site input and exportable comparison reports.
Patches to the Zephyr project: CMake and sysbuild fixes, NXP MCXN and MCXA board and devicetree support, the Renesas RA flash driver, USB CDC ACM logging, cbprintf bounds, UART async test stability and the OpenTitan PLIC.
Snake written in C for the TI Cortex-M3 Stellaris evaluation board. The first repository on this account.
The classic slot deck game in Objective-C on Cocoa.