anirudhux Hire me

Integrating Sarvam APIs: A Case Study

I use Sarvam on two sites. On anirudhux.com it voices the narrated summaries on eight case studies. Contraire runs a full Sarvam version at contraire.anirudhux.com/sarvam; the production Contraire stays on Claude.

Sarvam APIs
6
Audio sets
96
Total spend
₹1,339.30
API calls
~5k

Summary

Six APIs across the two properties. The Contraire fork reaches all of them; the portfolio uses two, Bulbul for voice and Mayura for translation.

API Model Where it runs What it does
Chat and reasoning sarvam-105b Contraire fork nine council perspectives and three reports, in the document's language
Text to speech bulbul:v3 both Contraire's report audio; all 96 portfolio combinations
Speech to text saarika:v2.5 Contraire fork mic dictation and the sample voice note
Translation mayura:v1 both Contraire's Indic samples; the portfolio's five languages
Language ID text-lid Contraire fork detects the pasted language
Document AI sarvam-vision Contraire fork extracts text from a PDF
the two properties, side by side

Swapping the engine under Contraire

Contraire runs a draft past nine AI readers, each pushing back in its own voice, and folds their reads into a rewrite. In production the reads run on Claude Haiku 4.5; the fork runs them on Sarvam 105B, slower per read but cheaper, and pre-baked so no visitor waits on it.

Reader Haiku time 105B time Δ speed Haiku cost 105B cost
The Critic 6.4s 13.1s 2.0× ₹0.26 ₹0.16
The Feminist Theorist 5.6s 12.7s 2.3× ₹0.26 ₹0.18
The Empath 4.4s 16.2s 3.7× ₹0.22 ₹0.18
The Old Guard 5.9s 14.3s 2.4× ₹0.25 ₹0.16
The Pragmatist 5.1s 10.5s 2.1× ₹0.24 ₹0.16
The Advocate 5.6s 30.2s 5.4× ₹0.25 ₹0.14
The Mentor 5.6s 10.4s 1.9× ₹0.26 ₹0.16
The Realist 6.5s 13.8s 2.1× ₹0.27 ₹0.17
The Moralist 5.7s 12.0s 2.1× ₹0.26 ₹0.15
Council total 50.8s 133.2s 2.6× ₹2.27 ₹1.46
The nine readers running on Sarvam 105B

Sarvam 105B vs Sonnet

Three reports follow the council: Synthesis, Drift and Variance. In production they run on Claude Sonnet 4.6; the fork runs all three on 105B, in parallel.

Stage 105B time Sonnet time Δ 105B cost Sonnet cost
Synthesis 32.7s 24.8s 1.3× slower ₹0.32 ₹2.13
Drift analysis 22.4s 25.7s 1.1× faster ₹0.28 ₹2.13
Agent divergence 38.2s 46.3s 1.2× faster ₹0.37 ₹3.25
Total 93.3s 96.8s about even ₹0.98 ₹7.51

105B splits one 4,096-token budget between thinking and answering, unevenly by task.

Stage Thinking Answer Headroom
Synthesis ~1,648 ~727 1,721
Drift 0 732 3,364
Variance ~2,693 ~799 604

Bringing voice to Synthesis, Analysis & Variance reports

Each report and each of the nine perspectives has a Play button, read by Bulbul v3. The same pipeline made Contraire's video voice-overs, in Sunny and Shreya, the two of six Bulbul voices I shipped.

Every report and perspective plays back in Bulbul's voice

Rebuilding the input methods

The fork started with one way to add a draft, a text box. I grew it to five, each powered by a different Sarvam API.

Input Sarvam API What it does
Paste text text-lid detects the pasted language
Sample documents mayura:v1 loads a ready draft in English, Kannada, Hindi or Tamil
Sample PDF sarvam-vision extracts text from a PDF, layout-aware
Sample voice note saarika:v2.5 a baked clip, transcribed
Live mic dictation saarika:v2.5 transcribes speech in real time
the input methods

Audio-visual case studies, powered by Bulbul

Eight case studies, each with a spoken summary. Bulbul voices Sunny and Shreya; Mayura translates. English, Kannada, Marathi, Hindi, Tamil, Malayalam.

Clip Section Length
Constella AI 2:59
Contraire AI 2:53
Shape of portfolios AI 2:56
Design System (Ennoventure) Work 3:03
ExpressBuy (PhonePe) Work 3:06
PhonePe Cart Work 2:52
UPI International (PhonePe) Work 2:29
UPI Lite (PhonePe) Work 2:27
Total 22:45
the player on anirudhux.com

Cost breakdown

Model Dashboard quantity Dashboard spend Portfolio Contraire fork
Bulbul v3 (text to speech) 345.6K chars ₹1,036.67 291.4K chars · ₹874.13 54.2K chars · ₹162.54
Mayura v1 Formal (translation) 146.5K chars ₹292.91 99.0K chars · ₹198.05 47.5K chars · ₹94.86
Sarvam 105B (chat) 519.4K tokens ₹5.73 none ₹5.73
Saarika v2.5 (speech to text) 3.7 mins ₹1.83 none ₹1.83
Speech and audio (misc) 103 mins ₹2.16 none ₹2.16
Total ₹1,339.30 ₹1,072.18 ₹267.12
Sarvam dashboard, usage by model