CAI Technology
Book a demo
Menu ☰
iris In development

IRIS — Persistent AI Orchestrator

Bring-your-own-LLM, bring-your-own-DB, bring-your-own-Telegram. We adapt to your stack, not the other way.

Request a demo →
The problem

Agent frameworks (LangChain, AutoGPT, CrewAI) force a single model, single queue, single chat. Small ops want to use their own Claude/Postgres/Telegram without complete re-engineering.

How it works

  1. 1

    Send command in chat or via API: 'archive backups older than 90 days to NAS'.

  2. 2

    IRIS analyzes the request, decomposes into steps, and proposes the plan with estimated impact.

  3. 3

    Approve (inline buttons or CAI-AUTH push) or reject. For high risk (>1 server, prod change), approval is mandatory.

  4. 4

    After execution, short report with outcome + full audit. Iris remembers and learns from interactions.

Capabilities

Propose-then-act

Never action without explanation. For read-only (logs, metrics, status) it's instant. For any material change, you see the plan first.

Hierarchical reasoning

Orchestrator → Supervisor → Agent. Smart routing across a premium model for design, mid-tier model for execution and self-hosted model for polling. Optimal cost, high quality.

Bring-Your-Own-Client

Adapters for any external LLM provider (subscription, OpenAI-compatible API) or self-hosted on-prem model. Adding a new adapter: ~150 lines.

Full audit log

Every task, proposition, plan, approval, execution, outcome — in a relational database with timestamp + actor. For retrospective audit (why did we do X in March?), it's ready.

Multi-channel intake

Chat (inline), web dashboard, push notifications, REST API. Use the right channel for the context — chat on mobile, web on desktop, API for CI/CD.

Tech stack

  • ▸ Orchestrator persistent cu izolare per-agent
  • ▸ Audit + state pe DB relațională
  • ▸ Aprobare push prin CAI-AUTH
  • ▸ API REST + integrare chat

Evidence

  • ✓ Smart routing significantly reduces cost vs running a single premium model end-to-end
  • ✓ Inline chat approval in under 3 seconds
  • ✓ Bring-Your-Own-Client: works with your existing LLM (external or self-hosted)
  • ✓ Full audit of every action for retrospective investigations

FAQ

When does IRIS execute without approval? +
Read-only (status, logs, metrics, inventory queries) — instant. For any write (create/modify/delete), service start/stop, config change — approval mandatory. Policy configurable per persona / per environment.
Does it replace Ansible/Terraform? +
No, it uses them. IRIS is the conversational orchestration layer; Ansible/Terraform/shell scripts are the tools. For production-grade applications, Git-versioned scripts remain the base; IRIS invokes them correctly, with approval.
Sensitive data in prompts? +
In standard config, IRIS runs with a self-hosted on-prem LLM for anything containing PII. Only design and planning go through external providers. Configurable per persona — some organizations enforce 100% local.

We start with a 30-minute conversation.

Free AI-readiness audit for companies with 50+ employees. We reply within 24 hours.

Open contact form →