Projects
Research prototypes, open-source tools, and applied systems.
MemoryQC-Edge: Label-Efficient On-Device QC Agent for MRI Segmentation
Under domain shift a segmenter can miss a tumor confidently, so QC built on uncertainty inherits a blind spot. This agent learns a failure detector over a compact 29-dimensional signature of the segmenter's output, flags likely failures for clinician review, and recalibrates online from the cases that come back reviewed, without touching the segmenter's weights. Accepted at the MedAgent workshop (MICCAI 2026).
- AUROC 0.91 on held-out BraTS-Africa; reviewing the top-ranked 20% of slices surfaces 48% of all major failures.
- numpy-only runtime scores a slice in 88 µs on a 15 W Jetson Orin Nano with model state under 200 KB.
EvCorrupt-Bench: Runtime Corruption Detection for Event Cameras
A benchmark and monitor study for detecting sensor corruption in event-camera streams at runtime: six physically modeled failure modes (leak-event hot pixels, mains flicker, activity bursts, dead regions, readout congestion, polarity faults) injected as timed episodes into 42 recordings across four sensor families, with leakage-safe injection and false-alarm-budgeted evaluation. Under revision for resubmission.
- Grew out of a self-supervised SNN-GNN event denoiser (97% reconstruction accuracy), which now serves as label source and fallback action.
- Injection pipeline released as the open-source `evlab` package (pip install evlab): reproducible recipes, per-event ground truth, clock-quantum snapping.
SpikeGuard-RC: Spiking Reservoir Safety Monitoring (in progress)
An always-on spiking reservoir that maintains a compact temporal memory over windowed event-stream features and predicts safe / unsafe / uncertain, gating downstream perception: continue, abstain, or fall back to denoising. Targeting a computer-vision venue submission.
- Corruption-injected freeway benchmark with exact onsets for AUROC and time-to-detection; UAV streams as real-world transfer test.
- Compared against rule-based, MLP, GRU, LSTM, and non-spiking ESN baselines on detection quality, RAM, latency, and spike-rate energy proxy.
FedCW: What Compression Does an Edge Federation Actually Need?
A measured federated-learning study on real hardware instead of simulation: a desktop, Raspberry Pi 5, Raspberry Pi 4, and Jetson Orin Nano train a shared autoencoder over live Wi-Fi with per-exchange wall-time and byte accounting, on a cosmic-ray sensor-network (CosmicWatch / CREDO) workload. Manuscript prepared for ICAIIC 2027.
- Straggler economics: staleness-weighted asynchronous aggregation reaches 2.4× lower loss at equal wall clock on the most heterogeneous fleet, at a 5.5× higher steady-state wire rate.
- Compression ladder with a recovered floor: naive int8 is the only wire format that degrades beyond run-to-run variation; error-feedback wires (quantized delta, top-k) match fp32 at 73–81% fewer bytes.
- Links decide participation: under emulated LTE / LTE-M / NB-IoT, the compressed wire admits 1.6–3× more updates, and a server-side fp32 → fp16 → delta-int8 policy adapts per connection with no client-side logic.
SpikeAdapt-SC: Adaptive Spiking Semantic Communication over UAV Air-to-Ground Links
Spiking semantic communication for UAV edge inference: binary spike codes carry visual and radar features over a physics-based air-to-ground link (BPSK over Rician fading, validated against its BSC abstraction), and a pilot-driven closed loop jointly selects two training-free rate axes, spatial masking ρ and temporal truncation T′, under a hard per-image payload cap.
- Aerial visual (AID / RESISC45): capacity rules out LDPC-coded JPEG under a 10 mW budget for BER ≥ 0.15, while the binary spike code loses only ≈4 / 6 pp at BER 0.30 (−8.4 dB); the joint loop cuts mission payload to 35–38% at parity with its full-rate checkpoint over 10 seeds, beating single-axis, encoder-matched 1-bit CNN, and HARQ-style policies.
- Radar features (RADIal, frozen FFTRadNet front-end): 58.2 Mbit/frame of pyramid features shipped as 2.3 Mbit of spike maps; detection holds F_H = 0.945 at BER 0.30 where 8-bit features under rate-1/2 and 1/3 LDPC fail past their decoding thresholds; the closed loop reaches mission P̄ = 0.947 at 48.7% lower payload.
- DoA study: a relaxed Ziv–Zakai bound flags the low-SNR threshold the Cramér–Rao bound misses. Measured 10 Hz operation on a Jetson Orin Nano.
PokeAgent Speedrun
Open-source Pokémon Emerald speedrunning agent with 4 agent scaffolds (four-module, simple, ReAct, ClaudePlays) and a multiprocess emulator/agent architecture for stable long-horizon execution.
- Documented lightweight simple pipeline runs 3–5× faster and retains the last 20 actions for compact action history.
- Production-oriented execution paths including no-ocr mode and automated gameplay recording at 30 FPS.