Cruq is now a member of NVIDIA Inception, NVIDIA's program for AI startups. It is free and takes no equity. Here is what it provides and where it fits in our work.
What the program provides
NVIDIA lists these benefits for members:
- Free self-paced courses and discounted workshops from the NVIDIA Deep Learning Institute.
- The latest SDKs, model libraries and developer platforms, plus preferred pricing on some NVIDIA hardware and software.
- Free cloud credits from NVIDIA and its partners.
- Introductions to investors in NVIDIA's venture network, for eligible members, and events with startups, VCs and NVIDIA staff.
The full details, and how to apply, are on the NVIDIA Inception page.
Where it fits
Speech models for Indian languages. Our voice agents need to understand people who speak Hindi, Bengali, Tamil and other Indian languages, often on a phone line. We train our own speech models for this, called svara, with PyTorch and CUDA on NVIDIA GPUs. The first one, a Hindi version of Whisper small, trained in 52 minutes on a single RTX 3090 and cut the word error rate on the FLEURS Hindi test set from 0.815 to 0.476. It is public on Hugging Face.
What comes next. The next steps are speech recognition for more languages and text-to-speech, which need longer training runs than one rented card can handle. We also want to serve voices fast and cheaply enough for live phone calls. That is where NVIDIA's training, software and compute offers help most.
Self-hosted models. Some customers need their data to stay on their own hardware. Cruq lets a workspace connect any OpenAI-compatible endpoint, so an open model served on the customer's own NVIDIA GPUs works like any other provider.
Customers still choose their models
Cruq is model-agnostic. A workspace connects its own provider keys and picks the model for each agent. NVIDIA Inception supports our work. It does not decide what a customer runs.
