GeoStratum combines Earth science methodology with production-grade software and Local-First, Embedded Edge AI. We develop tools for field geologists, researchers, and students, engineered for complete offline autonomy, zero cloud dependency, and total data ownership in harsh field environments.

---

🔬 Machine Learning & Edge AI Research

Our machine learning research focuses strictly on edge deployment and offline geological analysis:

1. Sub-Byte & Full-Integer Quantization (W4A8 & INT8)

• Zero-FP32 Execution Graphs : Enforcing strict integer kernels without floating-point fallback buffers, preventing memory stalls on mobile NPUs (Qualcomm Hexagon HTP, Google Tensor TPU, MediaTek APU, Intel NPU).
• Per-Tensor Signed Asymmetry (TENSOR_QUANT8_ASYMM_SIGNED) : Custom classification heads eliminating hardware partition breaks on mobile drivers.
• Kernel-Fused Convolutions : Structural absorption of 4D padding operators into native SAME convolution kernels.

2. Hierarchical Computer Vision (21-Pass Multi-Scale Tiling / 21-MST)

• Solving the fundamental resolution trade-off in geology: capturing both macroscopic outcrop textures and microscopic crystalline grains simultaneously in real-time, backed by C++23 SIMD auto-vectorization (ARM NEON / AVX2).

---

📦 Key Hub Repositories & Models

• GeoStratum/quantification-MobileNetV5-V4
  Technical documentation, full W4A8/INT8 quantization pipeline, and physical on-device benchmarks for MobileNetV5-300M and MobileNetV4-Large across Qualcomm Hexagon NPU (Copilot+ ARM64) and NVIDIA RTX Blackwell architectures.

---

🛠️ The GeoStratum Software Ecosystem

AI-Powered Solutions:

• 🪨 Lithotheque : Digital field notebook for Android & Windows. 100% offline resilience, embedded LiteRT rock classification across 104 lithological classes, sub-millimeter scale metrology, and AES-256 local encryption (Google Play, 1k+ downloads, 4.8★).
• 🏛️ Stratalog (In Active Development) : Native Windows 11 desktop application (C# WinUI 3 & C++ DLL) for vector stratigraphic column design, borehole logging, and local multimodal AI assistance (Google Gemma 4 E2B Instruct via Qualcomm QAIRT, Intel OpenVINO, and LLamaSharp).

---