Felix Moser

Project

STELLA: An Open Platform for Building, Testing, and Deploying Safe Conversational AI Agents

Conversation is becoming a primary interface to AI. As agents move out of text boxes and into health, work, and everyday life, the way they hold a conversation decides whether people trust them, stay engaged, and act on what they say. This is more than producing fluent text. A capable agent has to perceive the person it is talking to, decide what to say and how to say it, stay safe and on-task, and do all of this in real time across a multi-turn interaction. Most current systems collapse this into a single monolithic model and hope it works. We treat it as an engineering problem with separable, testable parts, and we set the bar with the hardest conversations rather than the easiest. STELLA is our open-source platform for building exactly these agents. It separates the underlying platform, which handles orchestration, session state, perception, and safety, from the agent logic that governs behavior, and it exposes an SDK so new capabilities can be built and swapped in independently. Every decision the agent makes is logged and inspectable, which is what regulatory frameworks such as the EU AI Act require and what most commercial systems cannot offer. STELLA also supports agent-to-agent simulation, so a new component can be stress-tested at scale before it ever reaches a real user. The larger goal is a modular ecosystem of agent capabilities, covering perception, dialogue behavior, expression, and safety, that can be developed in isolation and composed into a working system. The unifying design principle is adaptive support that scales to the individual, and the way to get that right is to derive the requirements from the most demanding user. People with early-stage dementia need an agent that notices subtle confusion, paces itself, repairs misunderstandings, and stays safe under conditions where ordinary systems quietly fail. GRACE, our first deployment, delivers Cognitive Stimulation Therapy to exactly this population. Designing for it fixes the specification for every component in the platform, and a system that holds up there scales down cleanly to the easier interactions that make up most of the market. That is also why the platform generalizes so far. Because the engine is built against the hardest case, it drops directly into far larger ones. Think remote patient monitoring, employee wellbeing and coaching, mental health support at scale, eldercare and accessibility, education, and enterprise assistants that handle real customer interactions instead of scripted flows. What makes this defensible is exactly what the rest of the field skips. Regulation is making auditability mandatory rather than optional, so a platform with traceability and safety testing built in turns a compliance burden into a moat, and the simulation layer lets teams validate agents before deployment instead of discovering failures live. STELLA is positioned to be the trusted infrastructure layer underneath the next wave of conversational AI use cases, in the categories where getting the conversation wrong is too costly to ship a black box.