Novaryn

The company

Building the AI operations layer for businesses

Every company will have an AI workforce. Novaryn is building the platform that operates it — the layer between a business and the AI doing its work.

The problem

Your employees are the integration layer

Software promised to connect a business's systems. In practice a person still moves information between email, the CRM, a spreadsheet, a document, a database and three chat tools — and that person is the integration.

This work is invisible in every org chart. Nobody's title is "moves data between systems", yet it is where a large share of an operations team's week goes — reading, copying, checking, chasing, and typing the same thing into a second place. It is too varied for traditional automation, which needs the rules written down in advance, and too structured to need a person's judgement.

That gap is exactly the shape of what language models became good at, and it is why this is a company now rather than five years ago.

The solution

AI employees, not assistants

The unit is not a chat message. It is a completed business task — with the input, the decision, the tool used, the action taken, whether a person intervened, and what it cost.

  • AI Sales Employee

    Qualifies every lead before a rep opens it

    Never unattended — discounting, contractual commitments, outbound to a named-account list.

  • AI Support Employee

    Answers what it knows, escalates what it doesn't

    Never unattended — refunds, account deletion, public statements about an incident.

  • AI Finance Employee

    Reads every invoice, reconciles every ledger

    Never unattended — payment execution, vendor bank detail changes, ledger adjustments.

  • AI Operations Employee

    Keeps the systems in agreement

    Never unattended — bulk deletion, schema changes, anything touching production credentials.

  • AI Research Employee

    Reads everything, cites its sources

    Never unattended — publishing externally, representing a finding as a company position.

  • AI Reporting Employee

    The weekly report, written before Monday

    Never unattended — distribution to investors or regulators.

The platform

What is actually built

Rendered from the system itself rather than described. These are the tables the product runs on.

Costed processes
12
each with measured manual effort, automation potential and build difficulty
AI employees
6
roles with charters, owned processes and published limits
Live connectors
18
of a catalogue that shows beta and planned honestly
Lifecycle stages
11
one enum from first contact to managed operations

Underneath: a multi-tenant Postgres schema where isolation is enforced twice, an AI gateway that no agent bypasses, a resumable agent loop, a workflow engine, hybrid retrieval over the customer's own documents, and a risk policy that gates an action on what the action is rather than on how confident the model sounds.

The whole stack self-hosts. Nothing calls a third-party service, which is what makes it deployable inside businesses whose data cannot leave the building — insurance, healthcare, financial services.

The business

Services are the wedge. The platform is the company.

An audit is how a business discovers what its manual work costs. It is also how we learn which processes are worth automating, which systems they touch, and where a human has to stay in the loop.

  1. AI audit

    A business discovers what its repetitive work costs, from a model rather than a workshop.

    Paid, credited against the build

  2. Implementation

    The highest-return processes become deployed AI employees and workflows.

    Fixed scope, fixed price

  3. Platform

    The customer operates them: approvals, knowledge, costs, evaluations.

    Recurring

  4. Managed operations

    Novaryn runs what it built — monitoring, incidents, optimisation.

    Recurring, monthly

Each audit adds to the process library that scores the next one. That is the compounding asset: which tasks businesses automate, which fail, where humans intervene, and what a completed task actually costs. Customer data is never used to train models without explicit consent — the intelligence is in the aggregate structure, not the content.

Where this is

Early, and specific about it

Novaryn has no published customers and no revenue to report. Rather than dress that up, here is exactly what exists.

Built and running

  • The scoring engine, and the audit it produces
  • Multi-tenant schema with row-level security, forced on every table
  • AI gateway — routing, cross-provider failover, budgets enforced before the call
  • Resumable agent loop with human approval gates
  • Workflow engine with graph validation
  • Hybrid retrieval over customer documents, with citations
  • Customer console and internal admin
  • The whole stack self-hosted, with no external dependency

Not yet

  • Visual agent and workflow builders — the engines exist, the editors do not
  • OAuth flows for the connector catalogue
  • The evaluation runner
  • Billing
  • Paying customers

Company metrics are computed from database rows and shown internally. A figure with no rows behind it is reported as unmeasured rather than as zero, which is why none appear on this page.

Next

The fastest way to understand this is to run it

The audit uses the same scoring engine described on this page, and takes four minutes.

Run the AI auditRead the architecture