Open Source Projects

Gradion AI Projects
freeact		Lightweight, general-purpose agent that acts via code actions combining Python code, shell commands, and programmatic MCP tool calls, generated in a single inference pass. For programmatic MCP tool calling ("code mode"), it generates typed Python APIs from MCP server schemas and discovers task-relevant tools via agentic and semantic search, loading only what's needed into the context window. Saves successful code actions as reusable tools, supports subagent delegation and application-level approval, and runs locally as both a CLI tool and Python SDK.
ipybox		Unified execution environment for Python code, shell commands, and programmatic MCP tool calls, all combinable within a single code block executed in a stateful IPython kernel. Generates typed Python APIs from MCP server schemas ("code mode") and supports application-level approval of individual tool calls and shell commands during execution. Runs locally with optional lightweight sandboxing via Anthropic's sandbox-runtime, and is available as a Python SDK, MCP server, and Claude Code plugin.
Personal Projects
bayesian-machine-learning		A collection of notebooks about Bayesian methods for machine learning, like Bayesian regression, Gaussian processes, Bayesian optimization, variational inference in Bayesian neural networks, ..., etc. Each notebook covers a single topic and combines an introduction, mathematical basics and a simple implementation.
perceiver-io		The perceiver-io library is a modular implementation of Perceiver, Perceiver IO, and Perceiver AR in PyTorch, with a PyTorch Lightning integration for distributed training and a Hugging Face integration for inference. The project provides both ported official models and custom models used in training examples.
super-resolution		This project provides a Tensorflow 2.x based implementation of three popular single image super-resolution models: EDSR, WDSR and SRGAN. Pre-trained weights, training and inference examples as well as a data loader for the DIV2K dataset are included.
fairseq-image-captioning		Implements an Image Captioning Transformer with the fairseq sequence modelling toolkit by combining ideas from Self-critical Sequence Training for Image Captioning and Bottom-Up and Top-Down Attention for Image Captioning and Visual Question Answering with the Transformer architecture.
face-recognition		Demonstrates how to build a face recognition system with Keras, Dlib and OpenCV. The process involves preprocessing images for face alignment, generating 128-dimensional face embeddings with a convolutional neural network (CNN), training classifiers on labeled embeddings and predicting identities of new inputs.
streamz		Streamz is a combinator library designed to integrate Functional Streams for Scala (FS2), Akka Streams, and Apache Camel endpoints, allowing seamless interoperability between these technologies. Camel endpoints can be integrated into FS2 applications with the Camel DSL for FS2 and into Akka Streams applications with the Camel DSL for Akka Streams.
akka-stream-eventsourcing		This project brings to Akka Streams what Akka Persistence brings to Akka Actors: persistence via event sourcing. In the same way as persistent actors in Akka Persistence, stateful request processors in akka-stream-eventsourcing form a consistency boundary around internal state but additionally provide type safety and end-to-end back-pressure for the whole event sourcing message flow.

Project Contributions
sandbox-runtime	contributions	The Sandbox Runtime is a lightweight OS-level sandboxing tool that restricts process access to the filesystem and network without requiring containers. It uses native OS mechanisms (sandbox-exec on macOS, bubblewrap on Linux) to enforce security boundaries, supporting network isolation, filesystem restrictions, and violation monitoring. Originally designed for Claude Code to enable safer AI agents.
llama.cpp	contributions	llama.cpp enables large language model (LLM) inference in C/C++, with minimal setup and state-of-the-art performance on a wide variety of hardware. It also supports integer quantization from 1.5-bit to 8-bit, which significantly enhances the speed of inference while reducing memory requirements.
eventuate¹	contributions	Eventuate is a toolkit for building applications composed of event-driven and event-sourced services that communicate via causally ordered event streams. Services can either be co-located on a single node or distributed up to global scale. Service state can be replicated with causal consistency guarantees and remains available for concurrent writes during network partitions.
akka	contributions	Akka is an open-source toolkit for building highly concurrent, distributed, and resilient message-driven applications, implementing the Actor Model on the JVM. It provides developers with tools to create systems that scale efficiently, self-heal, maintain responsiveness during failures, and handle high-performance scenarios, all while offering features like distributed data management, reactive streaming, and elastic deployment. My main contribution to Akka is the Akka Persistence module.
alpakka	contributions	The Alpakka project is an initiative for implementing reactive and stream-aware integration pipelines based on Akka Streams. It provides a domain-specific language (DSL) for reactive and stream-oriented programming, with built-in support for end-to-end backpressure. It supports a variety of integrations and ensures interoperability with other Reactive Streams implementations.
akka-persistence-cassandra¹	contributions	This project provides a persistence plugin for Akka Persistence that uses Apache Cassandra as storage backend. Akka Persistence enables stateful Akka actors to persist their state via event sourcing, using persistence plugins that implement a common interface for storage backends.
eventsourced¹	contributions	Eventsourced is a library for building reliable, scalable, and distributed event-sourced applications. It adds scalable actor state persistence and at-least-once message delivery guarantees to Akka, allowing stateful actors to persist received messages to append-only storage and to recover state by replaying received messages. It is now superseded by the Akka Persistence module of the Akka project.
ipf¹	contributions	The Open eHealth Integration Platform (IPF) offers a comprehensive set of tools for creating healthcare-related integration solutions. It is built on top of Apache Camel and provides a custom domain-specific language (DSL) for implementing enterprise integration patterns in healthcare-specific integration solutions, based on HL7 and IHE standards.
camel	contributions	Apache Camel is an integration framework that simplifies the integration of different systems by supporting numerous enterprise integration patterns and data formats. It can be used standalone or embedded in other applications, making it deployable to various environments including cloud-based services.

¹ I'm the project founder, sponsored by a client or employer.