Taruen

Fine-tune Automatic Speech Recognition models (e.g. OpenAI Whisper, Conformer) and Text-to-Speech (TTS) models for specialized domains, distinct accents, or low-resource languages. Experience includes reaching 28.5% WER on held-out Laz test sets, building Coqui STT models, and curating speech synthesis datasets (Chuvash TTS).

Convert PyTorch and TensorFlow models to ONNX format. Apply INT8/FP16 quantization, graph pruning, and operator fusion with Hugging Face Optimum. Benchmark and optimize models for specific target production GPUs to multiply requests/second and reduce cloud inference cost.

Deploy production models to Microsoft Azure Machine Learning (Azure ML) endpoints ensuring high availability and zero downtime. Load-test endpoints under heavy simulated concurrency using Locust to right-size compute and validate latency SLAs. Audit and harden Docker images in Azure Container Registry (ACR) to remediate known CVE vulnerabilities.

Architect Elasticsearch mappings, tokenizers, and custom analyzers for complex similarity queries (phonetic matching, orthographic homoglyphs, and multilingual cross-search). Build high-throughput ingestion and data clearance pipelines capable of populating tens of millions of records (such as 20M+ USPTO public trademark archives).

Engineer deterministic, rule-based NLP tools when training data is sparse or unavailable. Build finite-state morphological transducers (HFST, Apertium), syntactic and dependency parsers using Constraint Grammar (VISL CG-3) under Universal Dependencies, and rule-based machine translation engines.