| --- |
| title: PyIntel Research |
| emoji: 👁️ |
| colorFrom: indigo |
| colorTo: blue |
| sdk: static |
| pinned: false |
| --- |
| |
| <div align="center"> |
|
|
| # 👁️ PyIntel Research |
| ### *Pioneering Open-Weight Multimodal Edge Intelligence, Spatial Reasoning & Situated Perception* |
|
|
| [](https://huggingface.co/datasets/pyintel/vantage-spatial-pov-mix) |
| [](https://opensource.org/licenses/Apache-2.0) |
|
|
| </div> |
|
|
| --- |
|
|
| ## 🧭 About PyIntel |
| **PyIntel** is an independent open-source AI collective focused on building high-density, compute-efficient foundation models, datasets, and perception architectures for **embodied robotics, edge devices, and ambient assistants**. |
|
|
| Rather than scaling parameters blindly, PyIntel focuses on **data-centric intelligence, 3D geometric grounding, and architectural efficiency**—enabling sub-3B parameter models to reason like frontier systems. |
|
|
| --- |
|
|
| ## 🔬 Core Research Pillars |
|
|
| ### 1. 📐 Situated Spatial Reasoning & 3D Grounding |
| Teaching vision-language models to ground physical entities in 3D coordinate space, compute line-of-sight raycasts, and understand depth, clearances, and bounding geometries. |
|
|
| ### 2. 👤 Perspective-Taking & Theory of Mind (POV of Others) |
| Breaking free of the "mirror reversal" bug. Training models to translate fluidly between **egocentric** (first-person) and **allocentric** (third-person/observer) reference frames, and accurately model what other agents in the environment can or cannot see. |
|
|
| ### 3. 🧠 Episodic Soft Memory Projections |
| Developing lightweight projection bridges that compress historical video frames and sensor audio into dense **soft memory tokens**—giving edge models long-horizon memory without context window exhaustion. |
|
|
| ### 4. ⚡ Real-Time Duplex Interaction |
| Architecting low-latency, streaming vision-and-voice pipelines for interactive physical assistants ("Jarvis") running on consumer edge hardware. |
|
|
| --- |
|
|
| ## 📦 Flagship Artifacts & Releases |
|
|
| ### 📊 Datasets |
| * **[`pyintel/vantage-spatial-pov-mix`](https://huggingface.co/datasets/pyintel/vantage-spatial-pov-mix)** |
| *A 7,712-pair harmonized multimodal instruction dataset featuring situated 3D spatial queries, camera perspective translations, and Theory of Mind reasoning with native `<|channel>thought` chains.* |
| * **[`pyintel/open-board-registry`](https://huggingface.co/datasets/pyintel/open-board-registry)** |
| *Ground-truth structured specifications for 1,700+ microcontroller boards and embedded development platforms (pinouts, voltages, RAM/ROM limits, and communication interfaces).* |
|
|
| ### 🤖 Models & Architectures *(In Development)* |
| * **`PyIntel Vantage-E2B`** |
| *A unified multimodal spatial reasoning engine combining Google DeepMind's `EmbeddingGemma-2` (740M) and `Gemma-4-E2B-it` (2.3B) via an episodic Soft Memory Projection Bridge.* |
|
|
| --- |
|
|
| ## 🌐 Connect & Collaborate |
|
|
| * **Organization Hub:** [huggingface.co/pyintel](https://huggingface.co/pyintel) |
|
|
