Gradion AI

We build ML systems and AI agents for production. Drawing on deep AI engineering experience [1][2][3] and 20+ years of production software development [1], we move from idea to validated prototype quickly, adopt new AI capabilities responsibly, and ship systems that run reliably at scale. We also embrace vibe coding when it makes sense, but we are just as serious about the engineering foundations.

ipybox[↗], a Python code execution sandbox with first-class support for programmatic MCP tool calling. Generates typed Python APIs from MCP server schemas and executes code in a stateful IPython kernel. Features programmatic tool call approval workflows and lightweight sandboxing via Anthropic's sandbox-runtime.

freeact[↗], a lightweight, general-purpose agent that acts via code actions rather than JSON tool calls. Writes executable Python code capable of calling multiple tools, processing intermediate results, and using loops and conditionals in a single inference pass. Can create new tools from successful code actions, progressively building its tool library.

group-sense[↗], a library for detecting patterns in group chat message streams and transforming them into self-contained queries for downstream AI systems. Enables single-user AI agents to engage in group conversations using configurable criteria, eliminating the need for specialized multi-party conversation training.

group-genie[↗], multi-party conversation intelligence for AI agents. Combines group-sense's intelligent pattern detection with a flexible agent integration layer, allowing single-user agents to join group chats without modification. Provides technology-agnostic agent support with default implementations for Pydantic AI and OpenAI's Agents SDK.

hybrid-groups[↗], a platform that integrates group-genie into Slack and GitHub, allowing single-user AI agents to participate in group conversations without modification. Agents act on behalf of individual group members using user-specific credentials, enabling secure access to private resources.

Canto[↗]. Development of an AI-powered visual and hybrid search platform built as a cloud-native solution that scales seamlessly to thousands of enterprise customers worldwide. Training of custom query processing models with synthetic data to deliver highly accurate search results across diverse content types.

MerlinOne[↗]. Conception and development of a multimodal agentic AI search engine with semantic understanding of images, videos, documents, and audio, enhanced through domain-specific model fine-tuning for media organizations. In addition to many other production deployments, it powers the Associated Press Newsroom, and contributed to MerlinOne's successful acquisition by Canto in 2023.

Red Bull Media House[↗]. Global distribution of the in-house DAM system to support low-latency local access to content and write-availability under network partitions. Release of the underlying algorithms and protocols as open source project. Content playout platform for Red Bull TV's global streaming service, delivering content to millions of users worldwide with consistent performance.

Martin Krasser[↗]. ML and AI engineer with strong background in deep learning, agentic systems, distributed systems, and system integration. Specialized in developing reliable AI systems and operating them at scale in production environments. Extensive industry experience in both technical and leadership roles. Active open source contributor.

[1] Industry projects
[2] Personal projects
[3] Citations