Cloud Security Web LLC – Associate Software Engineer
Mar 2026 – Aug 2026Remote
- Optimized speech inference with ONNX Runtime on CPU, reducing latency from 2,409 ms to 435 ms (∼5.5×); optimized SymSpell post-processing from 2,330 ms to 282 ms.
- Exported Parakeet TDT 0.6B to ONNX and ran it with ONNX Runtime on CPU, comparing against the PyTorch baseline.
- Built a reusable evaluation framework for speech models covering WER, medical entity accuracy, quantization, latency, and regression analysis.
- Benchmarked Parakeet, MedASR, and Nemotron; Parakeet TDT v2 Q8_0 reached 14.03% medical WER and 1.59% LibriSpeech WER.
- Implemented speaker diarization post-processing and Dockerized async workers on Azure Container Apps and Service Bus.