01Evaluation
What counts as good real-time multimodal interaction?
Benchmarks need to measure timing, interruption, backchanneling, visual attention, repair, and human-state change, not just final answer quality.
- latency
- timing
- visual context
- repair
Research, test, and launch relationship-based AI.
We study how AI should interact with humans in long-term relationships, provide a live experimental platform and APIs, and turn validated research into real products.
- Experimental Platform — Live
- APIs — Available
Niro Fieldfield.nirolab.fun
In Session- EMOTION 12 · STABLE
- ENGAGEMENT 61
- VISUAL FOCUS 48
- DIALOGUE FOCUS 60
- DISTRACTION 38
// 01 — THE LOOP
A closed loop from research to product — and back.
Niro Lab is not a research institute with a product on the side, or a product company with a research blog. The loop is the lab.
// The output of research
The Relationship Protocol
What our research ultimately produces: a living specification of how an AI behaves inside one specific relationship. Built around dynamic memory and context, it holds the relationship's definition, stages, value function, and experiment-derived initial settings. It self-optimizes in use — and everything in it, down to voice and avatar expression, is offered as callable APIs.
{ } Everything in the protocol — including the expression layer — is exposed as callable APIsView APIs →
- Decision Layer · Core
- Dynamic Memory & Context
- Relationship Definition
- Value Function
- Initial Settings
- Relationship Stages
- ⟳ Self-optimizing
// 02 — OPEN QUESTIONS
Open questions for real-time human-AI relationships.
The hard part is not only making AI respond faster. It is learning when, why, and how an AI should enter an ongoing relationship across voice, vision, memory, and time.
- 02Safety
How does safety work while the interaction is still unfolding?
Real-time AI has to detect risky requests, boundary pressure, dependency, manipulation, and emotional escalation across speech, camera, and behavior, then respond in time.
- consent
- boundaries
- risk state
- response timing
- 03Explanation
How should AI explain what it sees, remembers, and decides?
A relationship system should expose uncertainty, state curves, memory use, and decision reasons through interfaces people can actually understand and contest.
- state curves
- decision logs
- memory views
- generative UI
- 04Steering
How can people steer long-horizon agents without leaving the relationship?
Users need live updates, pause and redirect controls, memory inspection, and recovery mechanisms that keep long-running AI behavior aligned over days, weeks, and months.
- live steering
- pause
- memory control
- relationship repair
Niro Lab studies these questions as live relationship protocols: each question becomes an experiment, each experiment changes the protocol, and each protocol can become a product only after validation.
// 03 — LIVE PLATFORM
A live experimental platform for human-AI relationship research.
Create AI roles, define relationship rules, run real-time voice interactions, track human states, and export capabilities through APIs.
- 01
AI Role Creation
Create AI roles with identity, relationship context, and goals.
Define identity, relationship, interaction rules, long-term memory, and a value function for each role.
- 02
Relationship Protocols
Define interaction rules, memory policies, and boundaries.
Specify how each relationship interacts, remembers, sets boundaries, and repairs itself.
- 03
Real-time Voice
End-to-end real-time voice interaction.
Low-latency end-to-end voice sessions with custom voices.
- 04
Generative Avatar
Real-time driven generative avatars.
Drive generative avatars in real time, with customizable characters.
- 05
Human State Tracking
Estimate human state curves from camera, voice, and behavior.
Track signals like attention, confusion, frustration, and emotional intensity as state curves.
- 06
APIs & Visualization
Export core capabilities through APIs.
Export roles, voice, avatars, state tracking, and data visualization via APIs.
The platform is designed to test how AI behaves in different relationships — not just how AI responds in one conversation.
// 04 — API
APIs: platform capabilities, offered as services.
The platform is what we run experiments on. The APIs are what you build with — each category exposes a specific, documented set of capabilities.
Role & Relationship
Role API
- Create and configure AI roles: identity, persona, goals
- Bind custom voices and generative avatars to a role
Relationship API
- Define relationship type, stages, and boundaries
- Set the value function for a specific relationship
- Load experiment-derived initial settings
Interaction Rules API
- Configure strategies: encourage / challenge / pause / explain
- Switch strategies in real time based on human state
Perception & Memory
Human State Perception API
- Estimate attention, confusion, frustration, and emotional intensity from camera, voice, and behavioral signals
- Confidence scores attached to every estimate
State Tracking API
- Session state curves, event timelines, decision records
- Query state history across sessions
Memory API
- Long-term memory read / write with per-relationship policies
- User-controlled inspection and deletion
Real-time & Visualization
Real-time Voice API
- Low-latency end-to-end voice sessions
- Custom voice creation and switching
Avatar API
- Real-time driven generative avatars
- Character customization and lip-sync
Visualization API
- Render state curves and session dashboards
- Export experiment data for analysis
// 05 — RESEARCH & PRODUCTS
Each direction: from question to product.
Every relationship we study starts as a research question, runs through the live platform, and ships as a product only after validation.
Future Self
Live Product — SomewhenHow should a possible future self encourage, challenge, and guide the present self?
Somewhen — a real-time voice and generative avatar product where you meet and talk with a possible future version of yourself. Somewhen.me
In Research AI Teacher
How should an AI teacher balance learning outcomes, confidence, discipline, and long-term trust?
Running as live experiments on the platform; no product commitment yet.
Awaiting Further Research Virtual Partner
How should an AI partner express intimacy while maintaining healthy boundaries?
Studied before any launch — with focus on intimacy, dependency, memory, consent, and healthy boundaries.
Awaiting Launch AI Companion / Coach
How should AI support, motivate, and adapt to a person over time?
Will be released only after further research and validation through the platform.
Partner Products
Researchers, universities, and companies can also build and launch relationship-based AI products on Niro Lab.
// 06 — PARTNERS
Build long-term human-AI relationships with Niro Lab.
For Researchers
- HCI
- Affective computing
- Long-term companionship
- Education technology
- Human-AI relationship research
- AI ethics
- Virtual character research
For Companies
- AI Teacher
- AI Companion
- AI Coach
- AI Sales Agent
- AI Advisor
- Generative Avatar Products
- Virtual Character Products
Ethics as Infrastructure
Relationship AI touches intimacy, memory, and trust. These principles apply to everything that runs on the platform — ours and our partners'.
- Consent First
- User-Controlled Memory
- Interpretable Signals
- Healthy Boundaries
Study, test, and build relationship-based AI with Niro Lab.
Request access
Tell us who you are — we reply to every request about the platform, APIs, or partnership.
Contact Us
NIROLAB.FUN