ivo's home

This client runs the DeepSeek Harness engine on Android 8.0+ via PRoot without needing Root or Termux. It can execute locally on your phone or connect remotely to a server via SSH. The interface supports chat, tool approval, and mobile screen control in both native and web modes.

Useful if you want a dedicated mobile frontend for DSH without managing a separate server. The main catch is that local execution requires about 1 GB of storage and the APK is not on the Play Store. It is not for users expecting offline model inference, as requests still go to your configured provider.

more: https://github.com/ccch1mneyyy/dsh-Go

my other side: ivo.blogspot.com

IBM Research released Palette, a methodology to fine-tune a 20B open-weight model for generating slide decks. It uses a LoRA adapter to separate editorial planning from geometric coding, allowing parallel slide generation. The system is trained on approximately 10,000 synthetic slides using the IBM Design Language.

This is useful if you need self-hosted, brand-compliant deck generation without sending confidential data to frontier APIs. The catch is the heavy reliance on synthetic data and a specific design language, which limits immediate generalization. You must be comfortable building a verification harness to catch layout errors in the generated code.

more: https://huggingface.co/blog/ibm-research/palette-deck-building

my other side: ivo.blogspot.com

InsightFace 2.0 provides a Docker-based server for local face detection and liveness checks that runs on a single Linux CPU or NVIDIA GPU. It supports INT8 quantization for searching over 50 million images on an RTX 5090, acting as a self-hosted alternative to AWS Rekognition. The Python SDK includes a CLI and GUI for private video processing with blur and mosaic tools.

This is useful if you need on-premises identity verification without sending data to third-party APIs. The main catch is that the pre-trained models are licensed for non-commercial research only, requiring separate contact for commercial deployment. It is not a drop-in replacement for general image processing but a specialized tool for biometric data.

more: https://github.com/deepinsight/insightface

my other side: ivo.blogspot.com

This project runs TensorFlow Lite on an ESP32-CAM to read physical water, gas, or electricity meters. It captures images via an integrated camera, processes them locally to extract digit values, and exposes the data via MQTT, REST API, or InfluxDB. The hardware cost is under ten euros, roughly the size of a small matchbox, and it supports OTA updates.

It is a practical solution if you have analog meters and want local data without cloud dependencies. The catch is that it requires specific camera hardware and physical installation close to the meter face. It is not a ready-made consumer plug-and-play device but a firmware project requiring some tinkering to set up.

more: https://github.com/jomjol/AI-on-the-edge-device

my other side: ivo.blogspot.com

ONNX Runtime is an open-source accelerator library designed to run machine learning models efficiently on local hardware. It supports models from PyTorch, TensorFlow, scikit-learn, and XGBoost, executing them across CPUs, GPUs, and specialized accelerators using graph optimizations. For inference, it replaces standard framework runtimes to reduce latency and lower server costs without changing model weights.

It is worth integrating if you are already deploying models in production and need to squeeze out performance on commodity hardware or edge devices. The catch is the friction of exporting your models to the ONNX format, which can fail silently or drop unsupported operators depending on your framework version. Skip this if you are still rapidly iterating on model architecture and changing layers daily.

more: https://github.com/microsoft/onnxruntime

my other side: ivo.blogspot.com

Floorplan is an open-source tool that uses Claude or Codex to generate interactive architectural maps of your codebase. It translates files, endpoints, and cross-repo web requests into plain-English diagrams you can zoom, query, and keep updated as you work. It also includes a health-check feature to catch common security oversights like exposed API keys or open database tables left behind by AI generators.

It is worth a look if you are juggling multi-repository apps built with AI assistants and need a quick way to verify what changed and how it connects. The catch is that it is tightly coupled to specific LLM workflows and outputs visual explanations rather than strict static analysis. Skip it if you prefer reading raw source files or need formal, standards-compliant UML documentation.

more: https://github.com/haydenschmitty/floorplan

my other side: ivo.blogspot.com

Hummingbot is an open-source framework for building and running market-making or arbitrage bots across centralized and decentralized exchanges. It provides an hbot command-line interface to manage strategy configurations, monitor trade execution, and handle API key encryption locally. You can deploy it via source install on Miniconda or through Docker containers to integrate with existing automated workflows.

It is a solid tool if you want to move beyond basic scripts and need a stable harness for connecting LLM agents to market data. However, it is not a set-and-forget trading solution and requires significant time to tune strategies and manage risk parameters. Expect a steep learning curve if you are not already comfortable managing exchange API connectivity and order book logic.

more: https://github.com/hummingbot/hummingbot

my other side: ivo.blogspot.com

MuJoCo is a C-based physics simulator for articulated structures now maintained by Google DeepMind. It provides precompiled binaries for Linux, Windows, and macOS, alongside official Python bindings and a JAX-accelerated implementation called MJX. The engine is designed for high-performance simulation of contact dynamics in robotics and biomechanics research.

If you are working on differentiable physics or reinforcement learning, this is currently the industry standard. However, the XML-based model definition format has a steep learning curve and feels archaic compared to modern game engines. It is a powerful tool for research, but overkill if you just need basic rigid-body physics for a general application.

more: https://github.com/google-deepmind/mujoco

my other side: ivo.blogspot.com

Upscayl is a FOSS desktop application that upscales low-resolution images locally using Real-ESRGAN models. It wraps a C++ Vulkan engine in an Electron frontend, processing files entirely on your own hardware rather than routing them through a cloud API. Packages are provided for Linux, macOS, and Windows.

It is worth keeping around if you need to bulk-enhance graphics or UI assets locally without paying per-image subscription fees. The main catch is hardware dependency: it requires a dedicated Vulkan-compatible GPU, and integrated graphics chips often fail to run the models.

more: https://github.com/upscayl/upscayl