Data Scientist (Speech)

Company Description

Smallest AI is one of the world’s fastest-growing speech AI research companies. We have some of the world’s best speech AI models and we are powering over 30,000 customers across 12 different countries to automate their calling operations. At heart, we are engineers, researchers, and a bunch of nerds who love diving deep into AI.

Role
You’ll work deeply at the intersection of model architecture, GPU optimization, and real-time speech systems. Expect to profile models on different GPU setups, optimize kernels, experiment with new architectures, and push the limits of speed + efficiency.

What You’ll Do

1. Dive into model architectures (ASR / TTS / SLMs) and optimize them for specific GPUs and hardware profiles

2. Build, debug, and tune kernels using CUDA / Tinygrad / AMD toolchains

3. Convert, optimize, and benchmark models using TensorRT, ONNX, and other inference engines

4. Work hands-on with PyTorch to train, fine-tune, and evaluate real-time speech models

5. Run large-scale experiments, manage datasets, and analyze model performance at scale

6. Productionize models for ultra-low latency speech workloads

7. Collaborate with research, infra, and product teams to push models into production

Requirements

• Strong experience with CUDA, Tinygrad, AMD GPU toolkit, or similar low-level GPU programming stacks

• Hands-on proficiency with PyTorch and Python

• Deep understanding of neural networks, training dynamics, and optimization

• Experience handling and processing large datasets

• Familiarity with production inference pipelines

• Strong problem-solving skills with ability to go deep into performance bottlenecks
  

Great to Have

• Experience training speech models (ASR, TTS, SSL, etc.)

• Familiarity with audio encoders, decoders, waveform models

• Experience with MLOps, experiment tracking, deployment pipelines

• Training or fine-tuning models for production / published papers

• Experience with TensorRT and ONNX Runtime