--- title: README emoji: 🌍 colorFrom: blue colorTo: indigo sdk: static pinned: false ---
GeoStratum Banner # GeoStratum ### Geoscience Software & Applied Edge AI Solutions **Digitizing the field. Centralizing your tools.** [Official Website](https://www.geostratum.eu) • [Lithotheque (Google Play)](https://play.google.com/store/apps/details?id=com.lithotheque.app.release) • [GitHub](https://github.com/GeoStratum) • [Contact Us](mailto:communication@geostratum.eu) ---
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](https://huggingface.co/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](https://www.geostratum.eu/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](https://www.geostratum.eu/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). ---
GeoStratum Round Logo GeoStratum © 2026 — Founded by Lohan GUYON • Varades, France [Website](https://www.geostratum.eu) • [LinkedIn](https://www.linkedin.com/company/geostratum) • [Instagram](https://www.instagram.com/geostratum_communication/) • [Bluesky](https://bsky.app/profile/geostratum.bsky.social)