2025 07 10 google ai future fund

Google AI Future Fund

dwani.ai now has Multi-modal Inference API capability. Along with the Android App, we provide self-hosting options using Open Weight models. 

  1. Funding - Yes, I am looking for seed grants upto USD 100k for developer salary and organisation expenses for 1 year. dwani.ai is bootstrapped and I have been working full-time since February 2025.

    • We require GPU's for dwani.ai inference - we are constrained by GPU availability. 

  2. Traction - Android App has 300 total installs, with 100+ active installs. 

    • We have conducted 3 workshops for students. Total 500 students were provided an introduction to get started building AI applications for Indian languages.  https://dwani.ai/dwani-ai-workshop.pdf .
    • We conduct one workshop per month at Universities. We do not charge the students/university , Entry is open to all.  For July we have 2 workshops planned.

  3. AI Models and Technology

    • https://github.com/dwani-ai/deploy  - contains the detailed steps for self-hosting dwani.ai. 
    • We are using Nvidia GH200/ arm64 on lambda.ai to host one instance of dwani.ai  
    • Models - Text + Vision : google/gemma-12B-IT, 
      • Text to Speech ; ai4bharat/IndicF5, onnx-community/Kokora-82M-v1.0-ONNX 
      • Speech Transcription - ai4bharat/indic-conformer-600m-mulitlingual, Systran/faster-whisper-larger-v3
      • Tool Call : qwen/qwen3-32B
    • What is working :
      • ASR/Speech transcription is great with AI4Bharat models for Indian languages. 
      • Text to Speech : IndicF5 is good for Indian languages, but it is compute intensive. 
      • Gemma3 works well for Text and Vision inference for English. Kannada and German.
    • What is on your wishlist ?
      • We would like tool_call to be improved in the next release of Gemma,
      • I am yet to explore in depth Gemma-3n with Matroshka Transformers for Multi-modal inputs. It had a few issues on release day .