Anvaya is an internal product built by Argmax Studio to solve a problem we repeatedly saw in organisations: important information exists everywhere, but reaching it is harder than it should be.

What started as an internal product has since been deployed across three organisations — Argmax Studio, a customer contact centre, and an ERP provider — with 100+ people using it every day.

Rather than replacing the systems organisations already depend on, Anvaya sits above them as a conversational access layer. Employees can ask questions in natural language and retrieve information from connected databases, documents, operational systems, and tools without needing to know exactly where the answer lives.

“It’s the first internal tool our ops leads open before they open Slack. That never happens.” — ERP Provider

Built as an internal product

Anvaya was not designed as a one-off implementation for a single client.

We built it at Argmax Studio as a reusable enterprise AI platform — one that could connect different organisations, data sources, permission models, and AI providers without rebuilding the core system each time.

Today, Anvaya is used across three distinct environments:

  • Argmax Studio — our own internal deployment.
  • Customer Contact Centre — supporting teams that work across operational information and customer-related workflows.
  • ERP Provider — giving operational teams a conversational interface over business information and connected systems.

Across these deployments, more than 100 users interact with Anvaya daily.

The deployments are different, but the underlying problem remains the same: organisations have information, but their people need a better way to access it.

Eleven sources, one search layer

Large organisations rarely keep their information in one place.

It sits across dashboards, databases, reports, internal documents, operational tools, and team conversations. Someone might know an answer exists somewhere, but reaching it usually means remembering which system to open, which terminology to use, and who to ask when the information is difficult to interpret.

Anvaya puts one conversational layer in front of those systems.

When someone asks a question, the assistant determines which sources are relevant, retrieves the required information, and combines the available context into a response. Citations allow the user to trace the answer back to its underlying sources.

The result is a single interaction model across systems that were never designed to share one.

Not every enterprise question has a single answer sitting inside one document.

A user might need to compare records, retrieve information from multiple systems, perform a calculation, identify a trend, and explain what it means.

Anvaya handles these requests as structured workflows:

Understand the question → Identify the required sources → Retrieve information → Perform authorised operations → Combine the results → Return a cited answer

This lets the user describe the outcome they need without having to manage every intermediate lookup themselves.

Answers rooted in company information

Anvaya uses retrieval-augmented generation (RAG) to connect language-model responses to the organisation’s own information.

Relevant information is retrieved before an answer is generated, allowing responses to be grounded in the sources connected to each deployment rather than relying solely on general model knowledge.

Citations provide an additional layer of traceability, allowing users to inspect the information behind an answer.

Capability with boundaries

An enterprise assistant needs to be useful without becoming an uncontrolled interface to internal systems.

Anvaya therefore exposes authorised database operations and predefined tools through controlled interfaces. The model can determine when a capability is useful, but access remains bounded by the permissions and tools explicitly made available to it.

That creates a clear distinction between:

  • What the assistant can understand.
  • What it can retrieve.
  • What it can calculate.
  • What it is authorised to execute.

Access shaped around roles

Different organisations — and different users within the same organisation — require different levels of access.

Anvaya uses a multi-tenant architecture with granular permissions, allowing capabilities and information access to be shaped around the user’s organisation, role, and authorised operations.

The same underlying platform can therefore support different deployments without giving every user the same access.

An AI layer that isn’t locked in

Anvaya is designed to remain model-provider agnostic.

Different AI providers and models can be configured behind the platform, allowing deployments to adapt as model capabilities, pricing, latency, and organisational requirements change.

The provider layer can evolve without requiring the surrounding permissions, retrieval workflows, APIs, or user experience to be rebuilt.

Built to feel quick

Enterprise questions can involve several retrieval and reasoning steps, but the experience still needs to feel responsive.

Anvaya uses caching to avoid repeating expensive retrieval and computation for frequently requested information and operations. Under measured production conditions, this helped the platform achieve a median cited-answer time below three seconds.

The current implementation also follows a zero-data-retention-oriented conversation workflow: conversation history is not persisted, while the architecture leaves room for carefully controlled history storage in future deployments.

What sits behind the experience

Several execution layers work together:

User question → Intent and task planning → Retrieval and tool selection → Permission checks → Data and knowledge access → Reasoning → Cited response

Infrastructure Layer Technology / Role
Storage & Vectors PostgreSQL & pgvector supporting knowledge storage and semantic search
Document / Operational Data MongoDB and connected organisational sources
API & Gateway FastAPI providing the service layer
Orchestration Python coordinating models, data sources, RAG pipelines, and approved tools
Client Layer Progressive Web App (PWA) for platform-agnostic enterprise access
Real-time Layer WebSockets for responsive interaction

The complexity stays behind the interface.

For the person using Anvaya, the experience remains simple: ask a question, get an answer, and see where it came from.

What the numbers show

Anvaya is now running across three organisations, including our own internal deployment, a customer contact centre, and an ERP provider.

More than 100 users interact with the platform daily across these deployments.

Under measured production conditions, the platform has achieved:

  • 11 organisational sources unified behind one conversational layer.
  • 100+ daily users across three deployments.
  • 94% of questions answered without escalation.
  • Median cited-answer time below 3 seconds.

The 94% and sub-3-second figures were measured across the ERP provider’s production question volume during a multi-week evaluation window on its live enterprise workload, rather than a synthetic benchmark.

The numbers matter because they represent actual use of the product rather than a demonstration environment.

Outcome

Anvaya has evolved from an internal Argmax Studio product into a reusable enterprise AI platform used across three organisations.

Its purpose remains straightforward: make existing organisational systems easier to use together.

Employees do not need to know which database contains an answer, which document holds the supporting detail, or which internal tool they should open first.

They ask.

Anvaya works out where to look, retrieves what it is authorised to access, performs the necessary steps, and returns an answer grounded in the organisation’s own information.

Replacing every internal system was never the goal.

Making those systems easier to use together was.

When information is spread across an organisation, the best search experience is often the one that understands the question before asking the user where to look.

Project highlights

  • Designed and built Anvaya as an internal Argmax Studio product in 14 weeks.
  • Expanded the platform across three organisational deployments.
  • Supports 100+ daily users across active deployments.
  • Unified 11 organisational sources behind one conversational search layer.
  • Implemented RAG workflows using PostgreSQL and pgvector.
  • Handled multi-step analytical questions and cross-source retrieval.
  • Added controlled access to authorised database operations and predefined tools.
  • Built a multi-tenant architecture with role-based permissions.
  • Supported configurable AI providers and models.
  • Added caching for frequently requested information and operations.
  • Followed a zero-data-retention-oriented conversation workflow.
  • Delivered the product as a Progressive Web App.
  • Achieved 94% of questions answered without escalation in the measured ERP deployment.
  • Delivered a median cited-answer time of under 3 seconds in the measured ERP deployment.