Tarteel Moves From AWS to CoreWeave, Cuts Costs by About 56%
Tarteel, an AI-powered Quran study app, has moved its entire cloud infrastructure from AWS to CoreWeave with the help of deployment partner Zeet. According to CoreWeave, the move brought a 22% improvement in latency and a cost reduction of about 56%.
- Tarteel moved its entire cloud infrastructure from AWS to CoreWeave using Zeet.
- According to CoreWeave, latency improved by 22% and costs fell by about 56%.
- Median request latency is 423.5ms at about 1,600 requests per minute on over 40 NVIDIA GPUs on average.
- The migration took 1-2 days in total.
- Tarteel uses NVIDIA NeMo and Riva with Triton and plans to expand into NLP and TTS.
What Tarteel does and why it switched
Tarteel is a voice-guided app that helps Muslims practice reciting the Quran. Users converse with it much like they would with a teacher. The app follows along as verses are recited, highlights words as they are said and flags mistakes in real time. Users can also search the Quran and related texts in Arabic by voice. It also detects incorrect word choices and points out the distinctions between similar verses.
According to CoreWeave, the app has 3M+ downloads and transcribes 250k+ audio hours per month. Tarteel was unsatisfied with legacy cloud offerings, citing a lack of observability for workloads, an underwhelming GPU selection and an unoptimized solution for inference and production workloads. CoreWeave says it offers out-of-the-box observability with access to Grafana and Prometheus, plus a native S3 solution that it says minimizes Tarteel's cloud and data transfer costs.
NVIDIA Riva and NeMo in production
Tarteel had tried various speech-to-text toolkits, but according to CEO and co-founder Anas Abou Allaban, many of them failed to meet both performance and accuracy requirements and were not optimized for production deployment.
On CoreWeave, the team can use NVIDIA's Riva and NeMo toolkits. According to the source, NeMo let the team quickly train and fine-tune ASR and TTS models on Tarteel's own data and custom hyperparameters. The team then used Riva to deploy the NeMo models with Triton. Riva also allows Tarteel to adjust production behavior through parameters such as audio chunk size, padding and Voice Activity Detection (VAD).
Support for scaling inference across multiple GPUs comes from NVIDIA Triton, CoreWeave says.
Results after the migration
CoreWeave reports the following figures for Tarteel after the move:
- About 1,600 requests per minute
- On average over 40 NVIDIA GPUs (A4000/5000s)
- 22% improvement in latency
- Median request latency of 423.5ms
- About 56% cost reduction
These figures come from CoreWeave's own account of the project.
The role of Zeet
Tarteel uses Zeet to deploy services to multiple cloud providers in multiple regions (US, EU, AP). According to CoreWeave, Zeet resolved issues such as load balancing and DNS setup, keeping total migration time to 1-2 days.
Abou Allaban said: "Zeet made the switch to CoreWeave a breeze. We had minimal downtime and did not need any K8s or CW expertise, we just clicked a few buttons and we were live." He also described the workflow: "Just write a Docker container and your app is ready to be shipped on Zeet and CoreWeave."
What it means and what comes next
The case is presented by CoreWeave as an example of a smaller AI company moving production inference away from a large legacy cloud provider, with a broad choice of NVIDIA GPUs and simple deployment tooling as the draws. Because the numbers are published by CoreWeave, they should be read as the company's own claims.
Tarteel plans to use NVIDIA's SDKs and toolkits for new use cases such as NLP and TTS. Its stated goals include semantic search, Q&A, forecasting and spaced repetition to find the best chapter to recite or revise, and digitizing Islamic and Arabic literature.
