Selected research and engineering projects, most recent first. For the narrative version — why these problems, and what connects them — see Research. For formal citations, see Publications.
An on-device fast path clears ~90% of flows in 4.78 ms (P99 6.16 ms) on a Raspberry Pi 4B; only ambiguous flows escalate to retrieval at the base station. +16% F1 over the strongest baseline under mobility-induced drift. IEEE EDGE 2026.
99,492 flows across 1,024 configurations from an AERPAW-calibrated Containernet digital twin. Dataset, simulator, and calibration data publicly released.
Recovering the set of attacks in a coordinated campaign from scenario-level labels — 0.88 set-precision — with decision rules auto-extracted from a surrogate tree and mapped to concrete operator mitigations. Under submission.
A testbed isolating one network factor at a time, used to show that trace augmentation is method-specific — the same augmentation that helps size-sequence classifiers hurts timestamp-based ones. IEEE IPCCC 2025.
Breaking the accuracy–scalability tradeoff in hybrid telemetry by replaying a near-switch window through every query, with FPGA compression to make replay affordable.
A graph neural network that predicts bug locations by combining code AST structure with GitHub issue text.