← Back to case studies

Confidential client · On-premise AI

A teacher's voice did not have to become a lost archive.

A spiritual organization held years of teachings, music, meditation guidance, and documents. After its Guruji passed away, devotees wanted a respectful way to find his teachings again - privately, in their own language, and with the original source close at hand.

Read the build story Discuss a private AI system
Open-source model stackNo public-cloud processing
On-premise AI workstation with RTX graphics card installed

Local system online

Archive processing stays on-site

The constraint that shaped everything

No recordings, transcripts, or devotional material needed to leave the organization's own environment.

5-10 TB

Private archive

Audio, texts, documents, bhajans, and teachings

RTX 5090

Local compute

32 GB VRAM workstation for on-premise AI work

One desktop app

Devotee access

Natural-language questions across the private server

The story

This was never just a chatbot project.

The real question was simple: how can a large and sensitive spiritual archive remain useful after the person at its centre is gone? The answer had to respect the material, the devotees, and the organization's need for privacy.

/01

Start with what exists

The archive was not a clean database. It was years of pravachans, bhajans, meditation guidance, texts, and documents in many formats.

/02

Make the voice usable

Recordings mixed Hindi with Punjabi, Sanskrit, English, and local dialects. Many included harmonium, flute, room noise, and overlapping voices.

/03

Make every answer traceable

We evolved the first RAG prototype into a retrieval-first experience: search the archive, show the source, then help the devotee understand it.

Developer workstation used to build and test the private archive system

The difficult work happened before the question box: cleaning data, carrying context forward, and making the archive searchable without losing its source.

The hard part

A recording is not automatically knowledge.

Off-the-shelf transcription is not enough when the speaker moves between languages, a bhajan blends voice and instruments, and the source quality changes from one recording to the next. We treated every transcript as a future source that someone should be able to check.

  • Organize and preserve source metadata, timestamps, and a link back to the original recording.
  • Improve speech focus when music, instruments, and room sound overlap with Guruji's voice.
  • Shape transcripts for retrieval, not only for reading: segments, context, and provenance matter.

From basic RAG to grounded answers

First find the teaching. Then help people understand it.

The first version was a basic retrieval-augmented generation system. Each iteration taught us something: a small local model should not decide whether to search; the system should search by default. The user should see the source and be able to correct the scope when needed.

A simpler, safer answer path

1Ask in natural language
2Search the private archive
3Rank the most relevant passages
4Answer with source context

Trust rule

If the archive does not provide enough evidence, the product should say so. It should not pretend to know.

GPU hardware selected for the on-premise AI workstation

Local-first infrastructure

The compute was selected for a practical reason: process sensitive material locally instead of sending it elsewhere.

Privacy was the product decision

A system that stays in the room.

For this archive, privacy was not a checkbox. It decided the deployment model, model choices, hardware budget, and operating rules. The system was designed to run locally with open-source components, on the organization's own private environment.

Local execution

Transcription, retrieval, and answering run on-site.

Open-source stack

The architecture avoids normal-operation dependence on paid model APIs.

Visible evidence

Source passages and metadata stay close to every answer.

Governed voice work

Voice preservation is permission-led, disclosed, and access-controlled.

Preserving a voice responsibly

The goal was remembrance, not imitation without limits.

Once cleaner speaker material was available, we also built a local voice-cloning platform. The work focused on preserving cadence and delivery for approved uses, while keeping the ethical boundary clear: generated voice must never be presented as a new original recording.

A clear governance baseline

Rights-holder approval, restricted use cases, clear disclosure, access controls, and auditable outputs should be part of every voice-preservation deployment.

Night-time local AI processing and application development workstation

Voice preservation layer

Built on carefully prepared, isolated material and intended for Foundation-approved experiences only.

Questions people ask

Simple answers to the technical questions.

What is a private RAG system?

A private RAG system searches an organization’s own material before answering a question. In this project, the archive, retrieval layer, and answer model run locally so the material does not need to be sent to a public AI service.

Why was transcription difficult for this archive?

The material combined Hindi, Punjabi, Sanskrit, English, and local dialects. It also included devotional music and overlapping voices, which makes it harder to isolate speech and create useful transcripts from long recordings.

Can an organization build an AI knowledge base without putting sensitive data in the cloud?

Yes. A local server can run transcription, embeddings, retrieval, reranking, and an answer model on-premise. The right design depends on the archive size, data sensitivity, expected users, and hardware budget.

How should voice cloning be handled for a deceased teacher?

It should be permission-led and governed. Rights holders should approve allowed use cases, generated audio should be disclosed clearly, and access to models, recordings, and outputs should be controlled and auditable.

Build a system people can trust

Sensitive knowledge deserves more than a generic chatbot.

We help organizations turn valuable private archives into practical, source-grounded systems - without giving up control of the material.

Map your private AI system