Voice AI / Space
    Submit
    Back

    egma

    View on GitHub
    Git Repo
    egma-ai

    An open source platform for simulation testing, production monitoring, and evaluating voice agents with custom caller personas.

    About egma

    Egma is an open-source platform designed for simulation testing, production monitoring, and continuous self-improvement of voice AI agents.

    For the Non-Technical Reader

    Think of Egma as a flight simulator combined with a black box recorder for voice bots. Before letting a voice agent answer real customer calls, Egma lets you test it against simulated callers who act angry, confused, or overly verbose. Once live, it records every conversation to pinpoint where the AI stumbled, ensuring your voice agent improves safely without risking real customer relationships.

    For the Technical Reader

    Egma provides end-to-end evaluation and observability tailored specifically for conversational voice architectures.

    • Simulated Regression Suites: Run synthetic voice and text multi-turn conversations directly from your codebase using the Egma CLI or coding agents.
    • Custom Graders & Personas: Define custom caller behaviors, evaluation rules, tool mocking, and pass/fail thresholds.
    • Integrations: Out-of-the-box support for LiveKit (JS/TS) and Retell AI frameworks.
    • BYOK Architecture: Bring Your Own Keys model ensures zero markup on LLM/TTS provider usage costs.
    • Deployment & License: Open-source under the MIT license, with options to self-host locally (default port 3101) or deploy via Egma Cloud.

    Why It Matters

    Voice AI applications present unique reliability challenges due to latency constraints, non-deterministic LLM tool calls, and unexpected audio inputs. By establishing an open-source evaluation standard with a transparent cost model, Egma bridges the gap between local agent development and enterprise-grade reliability. Furthermore, its roadmap towards automated pull requests based on production failure analysis marks a significant step toward self-healing voice systems.

    The "Voice AI Space Lab" Idea

    Create an automated "Chaos Caller Suite" that stress-tests emergency dispatch or drive-thru voice agents. By simulating callers with heavy background noise, rapid speech, mid-sentence interruptions, and abrupt mind-changes, developers can automatically evaluate latency impact and tool call robustness before pushing changes to production.

    Explore the project on the Egma GitHub Repository.