Published artifacts

Architectures built from scratch.
Parameters counted by hand.

per-cell argmax · 54×96
01 / EDGE DETECTION

ANetV1 — per-region UAV detection at the edge

Avneh Bhatia

17,392 parameters — 150× smaller than YOLO 134 img/s on an M-series GPU 17 KB at int8 — smaller than a JPEG
See it run
cl / α polar · re 1e5
02 / AEROSPACE

FoilCORE — in-context memory along airfoil polars

Avneh Singh Bhatia · Joshua Selvaraj

6,394 parameters beats public NeuralFoil checkpoints causal in-AoA-order polar sequencing
Read PDF
pitch + heave · flutter onset
03 / AEROELASTICITY

FlutterForm — learned aeroelastic flutter prediction

Exea Labs

predicts flutter speed v_f and frequency ω_f directly <2.7% error on withheld sections 50,000 aeroelastic training samples
Read PDF
learned directed graph · sparse
04 / BIOSIGNALS

SD-Former — sparse directed attention for physiological time series

Ankur Sharma

graph-conditioned biosignal architecture robust under subject & device shift continuous DAG learning as inductive bias
Read PDF Code
detection ∝ network topology
05 / SEISMOLOGY

EPGNN — multimodal seismic graph network for early signals

Arya Addagarla

detections propagate by topology, not thresholds local noise ≠ regional alert speed and spatial precision
Read PDF

20+ member proposals in flight.

Browse proposals →