Steerio | Devpost

Inspiration

Voice AI is getting incredibly advanced, but in regulated industries like finance or healthcare, one incorrect sentence can create serious compliance risk. That tension inspired Steerio. We wanted to make sure AI could speak in real time while still meeting strict policy standards. Instead of building a smarter voice agent, we focused on building a safer one.

What it does

Steerio is a real-time safety layer for AI voice agents. Before any response is spoken, an LLM judge evaluates it against compliance policies. If the response is safe, it is delivered. If it violates policy, it is blocked, modified, or escalated to a human operator instantly. A live dashboard lets operators monitor conversations, inject instructions, and intervene in real time. Full audit logs ensure compliance visibility.

How we built it

We built Steerio on LiveKit's real-time voice infrastructure with Python. The core is a wrapper around voice agents that intercepts LLM responses before text-to-speech conversion. We use Supabase for policy storage, WebSockets for the operator dashboard, and OpenRouter for LLM-based judge evaluation. The entire pipeline is optimized for sub-200ms latency to keep conversations natural.

Challenges we ran into

The biggest challenge was balancing speed and safety. Every compliance check adds delay, which can hurt the user experience. Defining what counts as unsafe was also difficult, since policies are often nuanced and context dependent. We had to carefully tune the system to avoid both overblocking and underblocking.

Accomplishments that we're proud of

We successfully built a real time compliance filter that does not noticeably slow down conversations. We also created a flexible policy framework that can adapt to different industries. Most importantly, we demonstrated that safety and performance can work together.

What we learned

We learned that building responsible AI requires more than just technical skill. It requires understanding risk, regulation, and user trust. We also learned that constraints around speed and compliance lead to better system design decisions.

What's next for Steerio

Next, we want to expand beyond LiveKit to support more voice platforms and improve multi-judge consensus mechanisms. We plan to add automatic policy learning from operator interventions and build analytics to track compliance trends over time. Our goal is to make Steerio a foundational safety layer for enterprise voice AI across any platform. The main issues were missing the dashboard/human oversight feature and claiming "risk scoring" as future work when it already exists.