Spaces:

---

smolagents

/

ml-intern

like 280

Running on CPU Upgrade

🧠 Aura-Agent: Neural Guardian

A proactive, mobile-first seizure prediction system with a 30-minute pre-ictal warning, built end-to-end with PyTorch and smolagents.

🔗 Live Demo: Babajaan/Aura-Agent-Neural-Guardian

What It Does

Aura-Agent predicts epileptic seizures 30 minutes before onset from raw 19-channel EEG signals and deploys to Android via an ultra-lightweight student model.

Architecture

Key Technical Highlights

1. RGF-Net Student Model (< 100K params, ~36.7 KB quantized)

• Ring-Buffer GRU enables O(1) constant-memory streaming inference on-device
• FiLM conditioning (Perez et al., 2018) modulates GRU features using patient biometrics: (1+γ)⊙h + β where γ,β are generated from [age, sex, HR, HRV, seizure_history, medication_adherence]
• EEGNet-inspired depthwise-separable convolutions for temporal+spatial feature extraction
• Supports ONNX export and INT8 dynamic quantization for Android Edge-AI

2. Three-Component Knowledge Distillation Loss

• Soft-label KD (Hinton, 2015): Temperature-scaled KL-divergence (T=4.0)
• Hard-label CE: Class-weighted cross-entropy for pre-ictal/inter-ictal imbalance
• Feature distillation (TimeKD, 2025): SmoothL1 alignment of projected student embeddings to teacher embedding space — critical for cross-architecture (Transformer→GRU) distillation

3. smolagents CodeAgent for Autonomous Emergency Response

• RGFNetSeizureRiskTool (Tool subclass with lazy setup()) — runs model inference
• send_sos_alert, get_patient_gps, dispatch_first_aid (@tool decorators)
• Agent follows a strict medical protocol: Assess Risk → if ≥ 0.75 → GPS → SOS → First-Aid
• CodeAgent chosen over ToolCallingAgent specifically for if/else conditional branching

Grounded in Literature

The architecture and training recipe draw from:

• STAN (arxiv 2511.01275) — Spatio-temporal attention for seizure forecasting
• FiLM (arxiv 1709.07871) — Feature-wise Linear Modulation
• Hinton KD (arxiv 1503.02531) — Knowledge Distillation
• TimeKD (arxiv 2505.02138) — Feature distillation for time-series
• Edge DL (arxiv 2012.00307) — Quantized models for neural implants

Interactive Demo Features

The Gradio Space includes 6 tabs:

1. 🏗️ Architecture — Inspect both models with full parameter breakdowns
2. 🎯 Train & Distill — Run the KD pipeline live with tunable hyperparameters
3. 🔮 Live Inference — Test RGF-Net with synthetic EEG and adjustable patient biometrics
4. 🤖 Agent Workflow — Visualize the smolagents CodeAgent architecture
5. 📐 KD Loss — Mathematical walkthrough of the distillation loss
6. 📁 Source Code — Project structure and references

Built with PyTorch, smolagents, and Gradio. Feedback welcome!