Argus

Argus was inspired by firsthand experience in regulated manufacturing. Last summer, I (Syon) worked as a chemical process engineering intern at Kedrion Biopharma, where I saw that the biggest risks were not chemistry or equipment failures, but execution and documentation errors. Operators were expected to follow hundreds of pages of frequently updated SOPs while working in layers of PPE, without the ability to reference manuals mid-task. Steps were logged after execution, with verification relying on human memory and several layers of supervision and quality assurance. Even with these controls, a single mistake could invalidate an entire batch (costing us millions), trigger regulatory action (potentially shutting down plants), or delay lifesaving therapies (lives can be lost). That experience raised a simple question: in the age of modern AGI what if physical execution itself could be verified, not just recorded?

To address this, we built Argus, a real-time execution governor for life-critical procedures. Rather than treating SOPs (standard operating procedures) as static instructions, Argus turns them into enforced, stateful workflows where progress depends on real-time evidence. Operators interact with Argus hands-free through voice, while the system continuously verifies that required actions actually occurred. Vision (Overshoot API) is used to observe and confirm physical actions in the workspace match that of the SOP database using RAG, hardware signals are used to validate machine conditions such as temperature, pressure, or RPM against SOP-defined specifications, and a deterministic state machine governs whether a step can advance or must be locked. LiveKit plays an essential role beyond supplying voice interaction by orchestrating the real-time integration of voice commands, vision events, and hardware signals into a unified control loop that keeps all components synchronized.

Building Argus taught us how challenging and important real-time, multimodal enforcement is in the regulated environments of healthcare and the life sciences and beyond. We learned how to coordinate agentic systems across vision, voice, and hardware while maintaining low latency, reliability, and determinism. Integrating hardware signals and tuning vision pipelines for stability took significant effort, and orchestrating everything through a real-time communication layer was nontrivial, especially under hackathon time constraints. However, these challenges reinforced the core insight behind Argus: in life-critical workflows, trust cannot be based on human memory or retrospective documentation. It must be built directly into execution itself, so that correctness is enforced as the work happens.