Research, models, architecture, and AI infrastructure.
Atina Labs is the technical home for how Atina is built: model work, benchmark decisions, embedded assistant architecture, organization AI, and the infrastructure choices behind reliable AI assistants.
ICL-1 brings intent routing to embedded Atina.
Our first model track focuses on the moment before generation: deciding whether an embedded assistant needs company knowledge, product data, live tools, or general reasoning.
ICL-1: Intelligent capability routing for embedded Atina agents
ICL-1 is Atina's first capability router for embedded agents: a lightweight MiniLM-based classifier that decides which customer capability should be activated before the main response is generated.
Read Featured ReportRecent decisions from the Atina engineering workspace.
Release note: Embedding, Organization AI, and ICL-1
The July 2026 Atina Labs content refresh documents embed setup, organization training, ICL-1 routing, embedded RAG architecture, infrastructure capacity, and a latency-based speech decision.
Atina AI assistant infrastructure
A public overview of Atina platform capacity: chat, embed, organization training, knowledge workflows, capability routing, usage controls, analytics, support, and avatar experiences.
Embedded agent RAG architecture
The upgraded Atina Embed architecture: domain context, ICL-1 routing, knowledge retrieval, product feed search, read-only live tools, escalation, response generation, and analytics.
Organization AI training with company handbooks
How organization teams use Atina to turn company handbooks and internal documents into private knowledge bases, training modules, lessons, learner support, access control, and reports.
Explore the systems behind Atina.
Find model reports, benchmark outcomes, architecture notes, embedding guidance, organization AI research, infrastructure updates, experiments, and release notes.
Release note: Embedding, Organization AI, and ICL-1
The July 2026 Atina Labs content refresh documents embed setup, organization training, ICL-1 routing, embedded RAG architecture, infrastructure capacity, and a latency-based speech decision.
Atina AI assistant infrastructure
A public overview of Atina platform capacity: chat, embed, organization training, knowledge workflows, capability routing, usage controls, analytics, support, and avatar experiences.
Embedded agent RAG architecture
The upgraded Atina Embed architecture: domain context, ICL-1 routing, knowledge retrieval, product feed search, read-only live tools, escalation, response generation, and analytics.
Organization AI training with company handbooks
How organization teams use Atina to turn company handbooks and internal documents into private knowledge bases, training modules, lessons, learner support, access control, and reports.
Embedding Atina on third-party sites
A practical guide to launching Atina Embed: choose the right plan, create an embed app, authorize domains, configure capabilities, copy the SDK snippet, and test before production.
Why Atina rejected the Gemini TTS orchestration hack
An architectural decision note explaining why Atina kept its current avatar response path instead of adding a second Gemini TTS pass that increased user-perceived latency.
ICL-1: Intelligent capability routing for embedded Atina agents
ICL-1 is Atina's first capability router for embedded agents: a lightweight MiniLM-based classifier that decides which customer capability should be activated before the main response is generated.
How Atina is expanding from assistant to infrastructure.
A documented drop-in snippet and playground for bringing Atina into external websites.
Company handbook intelligence for onboarding, training, mentoring, and knowledge work.
Domain-controlled assistants with knowledge sources, product feeds, and live tools.
A lightweight intent model for routing embedded assistant requests before generation.
Speech, vision, and reasoning research will enter Labs when they meet Atina’s quality and latency bar.
What the platform is learning from now.
Organization AI training with company handbooks
How organization teams use Atina to turn company handbooks and internal documents into private knowledge bases, training modules, lessons, learner support, access control, and reports.
Embedding Atina on third-party sites
A practical guide to launching Atina Embed: choose the right plan, create an embed app, authorize domains, configure capabilities, copy the SDK snippet, and test before production.
Why Atina rejected the Gemini TTS orchestration hack
An architectural decision note explaining why Atina kept its current avatar response path instead of adding a second Gemini TTS pass that increased user-perceived latency.
ICL-1: Intelligent capability routing for embedded Atina agents
ICL-1 is Atina's first capability router for embedded agents: a lightweight MiniLM-based classifier that decides which customer capability should be activated before the main response is generated.
