This is the lab of Emilio Girard (Montréal). One person, a DGX Spark, two Jetsons, and rented Blackwell cards when a run needs them. Site: emiliogirard.com.
Public work here:
| Repo | What it is |
|---|---|
| joules-per-task | Hardware energy measurements for learned policies on a Jetson Orin Nano: a VLA policy at 963 J per successful task cut to 489 J, plus a four-policy spectrum. Paper draft, harness, raw logs. |
| per-block-drafter-selection | 7,309 speculative-decoding blocks with four drafters verified on identical prefixes. Workload routing is a null; per-block selection captures 61% of the oracle gap at 43% of the cost. |
| risk-aware-terrain-jetson | IEEE T-IV 2026 terrain-navigation reproduction on the Orin with energy numbers. |
| cbcl-pr-jetson | Class-incremental learning reproduction, five seeds, INT8/FP16 ablation. |
| pylox-vision-demo | 8-camera detection and re-identification stack on Jetson. |
Models on Hugging Face: huggingface.co/pyloxsystems. Every card carries a benchmark table, including the three adapters that lose to their base model.
Contact: pyloxsystems@gmail.com
