From catalogue to cinema.

Lahari Music holds one of India's largest devotional catalogues: 1,500 songs in Sanskrit, Kannada, Tamil, Telugu and Malayalam. Only a fraction of it had ever been monetized.

We built a production stack that turns a song's audio into publishable content in every format, at a speed and cost that make the whole catalogue worth producing.

1,500

songs, five languages

97%

lyrics on the beat

800

lyric videos in two months

100+

music videos in six weeks

How we achieved this.

A catalogue worth more than it was earning

Every song in the catalogue could be a lyric video, a music video, a set of Shorts and subtitled versions in five scripts, with new content for every festival. A lyric video took an editor six hours, and a music video took two weeks and thousands of dollars. At that speed and cost, most songs would never earn back what they took to make. Only the biggest songs got made, and about a tenth of the catalogue reached YouTube. We set out to build a production stack that could cover every song. The first step was moving the catalogue to the cloud, with every song, lyric and metadata record accessible to agents.

We started with the lyrics.

Nearly every format depends on the lyrics, timed to the singer. Off-the-shelf speech recognition is built for speech, and it breaks on sung Sanskrit content. We built a proprietary transcription stack that managed to transcribe the entire catalog with 97% accuracy.

97% of lyrics land on the beat, against 55% with standard AI tools
Subtitles in Sanskrit, Kannada, Tamil, Telugu and Malayalam, all in sync with the singer

Eight hundred lyric videos

With timed lyrics in place, lyric videos became a production line. A separate pipeline turned Lahari's deity artwork into 1,500 animated clips, which loop behind the lyrics. Each video pairs the song with the right deity and the right script. It is rendered in the cloud and checked by a person before release.

The line soon ran smoothly enough to be driven from a chat. An agent in Telegram knows the catalogue, the render queue and the festival calendar. It suggests what to make ahead of each festival, takes edits and starts renders once someone approves, all over text messages.

800+ finished videos
About 10 minutes per video, down from six hours

Music videos were the hard part.

Music videos were harder. Video models are powerful, but they are expensive and difficult to direct. A music video needs 30 to 40 clips that belong together: the same faces and the same world from shot to shot. Assembling that by hand is slow, most generations miss, and getting it right takes skilled people watching every run.

So we built a custom video production harness around the models. It starts from the song, its lyrics and its mood, and writes the concept and the visual script. Characters, places and style are designed once and reused in every shot. Each shot is storyboarded and approved before any money goes into video. An AI director works through the shots alongside the artists. When one detail is off, they fix that detail instead of regenerating the whole shot.

About 1.5 attempts per usable shot, where the industry needs around 4
A finished music video in about a day, where AI-assisted teams take a week
Claude and Codex work in the studio with the same tools the artists use

That made music videos viable for the entire catalogue. Three artists made more than 100 in six weeks, and the new content has already drawn close to 500,000 views.

One song, many formats

Every song now yields a full set: a lyric video, a music video, Shorts and Reels, thumbnails, tags and subtitles in five languages. One production run builds a bank of content instead of a single upload.

Each song makes the next one easier

Everything the system learns stays with it: the characters, the places, the shots that worked and the label's preferences. Each new song starts from what the last one built. We have since turned it into a product, Sacred, so that other labels can do the same with their catalogues.