The September 21 awesome-infra-for-ai release lists seven projects in four sections: AI safety, model evaluation, model-serving infrastructure, and agent workflow orchestration.

Doberman-Core, a local action filter for coding agents, works between an agent and its tools and stops destructive commands and commands that exfiltrate data. Probity checks every file write and shell command before it runs to enforce development rules. OpenAFW acts as a local proxy in front of Anthropic, OpenAI, and Gemini API calls and replaces secrets in outgoing requests with placeholders.

The evaluation section includes OpenCompass for running standardized model benchmarks and Keystone Bench for testing changes to a chat assistant’s clinical advice. fastembed-rs creates embeddings, numerical representations of text and images, and reranking scores locally through ONNX Runtime, a model runtime, without Python or a GPU.

Idun turns a LangGraph or Google ADK agent into a FastAPI service with chat, an admin panel, tracing, and MCP tool governance.

Claim check:

  • The September 21 release lists seven projects in four sections: AI safety, model evaluation, model-serving infrastructure, and agent workflow orchestration. (confirmed by the publication itself: evidence; «AI Safety & Guardrails DobermanCore/Doberman-Core nizos/probity openguardrails/openafw LLM Evaluation & Testing open-compass/opencompass xinxuxin/keystone-bench Model Serving Frameworks anush008/fastembed-rs Workflow Orchestration for AI Idun-Group/idun-agent-platform»)
  • Doberman-Core, a local action filter for coding agents, works between an agent and its tools and stops destructive commands and commands that exfiltrate data. (confirmed by the publication itself: evidence; «Local, open-source runtime guardrail that sits between AI coding agents and their tools, giving every action a PASS/AUTH/BLOCK verdict to stop destructive or exfiltrating commands before they run.»)
  • Probity checks every file write and shell command before it runs to enforce development rules. (confirmed by the publication itself: evidence; «checking every file write and shell command before it runs.»)
  • OpenAFW acts as a local proxy in front of Anthropic, OpenAI, and Gemini API calls and replaces secrets in outgoing requests with placeholders. (confirmed by the publication itself: evidence; «Local proxy that sits in front of Anthropic, OpenAI, and Gemini API calls from coding agents, replacing secrets in outgoing requests with placeholders and restoring them in responses so raw keys never reach the model or provider.»)
  • The evaluation section includes OpenCompass for running standardized model benchmarks and Keystone Bench for testing changes to a chat assistant’s clinical advice. (confirmed by the publication itself: evidence; «Open-source LLM evaluation platform for running standardized benchmarks across models, with configurable datasets, prompt templates, and an official public leaderboard. xinxuxin/keystone-bench - Evaluation benchmark that tests whether chat assistants change clinical advice correctly when a decisive fact is added, removed, or contradicted»)
  • fastembed-rs creates embeddings, numerical representations of text and images, and reranking scores locally through ONNX Runtime, a model runtime, without Python or a GPU. (confirmed by the publication itself: evidence; «Rust library for generating text, sparse, and image embeddings and reranking scores locally via ONNX Runtime, without a Python or GPU dependency.»)
  • Idun turns a LangGraph or Google ADK agent into a FastAPI service with chat, an admin panel, tracing, and MCP tool governance. (confirmed by the publication itself: evidence; «Self-hosted, open-source production wrapper that turns a LangGraph or Google ADK agent into a FastAPI service with a chat UI, admin panel, tracing, guardrails, memory persistence, MCP tool governance, and prompt management.»)

Publications:

Primary sources:

score 54.1 out of 100 · kind: release