Research
Notes from the lab.
Technical notes, open problems, and empirical findings from the Fantasma Synergy research team.
Attention Is Not Enough: Toward Structured World Models at the Edge
Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.
Calibration-Free Gaze Tracking at 60fps on Commodity Hardware
Classical gaze tracking requires a 90-second calibration ritual that most users abandon. We trained a universal gaze model on 4.2M synthetic eye-renders that generalizes to unseen users with 0.4° accuracy — no calibration, no personal data.
Byzantine Consensus in Low-Power Edge Meshes: Revisiting PBFT for the Real World
Classical Byzantine fault tolerance assumes reliable networks and ample compute. Edge meshes have neither. We present a modified PBFT protocol that tolerates 30% Byzantine nodes with 2ms consensus on a 10-node mesh running at 0.8W per node.
Detecting Concept Drift in Production ML Without Ground Truth Labels
Ground truth labels lag by days or weeks in production systems. By then, your model has already been making bad decisions. We present a distribution-matching approach that detects covariate shift within 3 hours of onset, using only model inputs.
The Intent Graph: Representing User Intent Across Modalities
Gaze, gesture, voice, and biosignals each carry partial intent signals. Fusing them naively loses the temporal structure that makes them meaningful. We propose a graph-based representation that preserves modality relationships and enables 6ms intent resolution.
On the Thermodynamics of Intelligence: Power, Latency, and the Edge Decision Curve
Intelligence is bounded by physics. We derive a first-principles framework for reasoning about the power-latency-accuracy tradeoff space for on-device inference, and show that most edge AI systems operate at less than 4% of the theoretical efficiency frontier.
Seven Open Problems in Edge Intelligence
A non-exhaustive list of the problems we find genuinely hard — where the literature is thin, the intuitions are wrong, or the solution space hasn't been properly charted.
Latency Budget Allocation Across a Distributed Inference Pipeline
When your end-to-end SLA is 12ms and you have 6 stages, how do you allocate the budget? We formalize this as a constrained optimization problem and describe the scheduling policy we use in Neural Fabric.
Anomaly Detection in Distributed Climate Sensor Networks
Climate sensing networks generate terabytes daily. Traditional threshold-based anomaly detection misses 60% of real anomalies and generates 10× more false positives. We deployed Edge Lattice across 2,000 sensors and reduced false positives by 87%.
Surgical Intent Recognition: From Gesture to Anticipatory Assistance
A surgical robot that reacts is a power drill with Wi-Fi. A surgical robot that anticipates is a partner. We trained Human Signal on 800 hours of annotated laparoscopic procedures and achieved 91.3% intent prediction accuracy 400ms ahead of action onset.
Neural Fabric v2: What Changed and Why
The first version of Neural Fabric was built for homogeneous compute clusters. v2 had to work on a 4W ARM chip next to a 400W H100. This required rethinking the scheduler, the memory allocator, and the model sharding approach from scratch.
Signal Integrity in High-Frequency Financial Inference
Market microstructure creates a uniquely adversarial distribution shift problem. The signals that predicted returns last quarter actively cause market impact this quarter. We describe the feedback-aware training loop we built for a tier-1 trading desk.
When Real Data Is Impossible: Building Synthetic Sensor Distributions for Rare Events
You cannot collect enough real data for events that happen twice per decade. We built a physics-informed generative model for rare climate events and used it to train a detection system that would otherwise require 40 years of historical data.