Ankit Dhiman

Fine-tuned a Qwen3-0.6B-based TTS model, achieving a 4.5/5 MOS in blind evaluations. Built an efficient inference-serving pipeline with NVIDIA Triton Inference Server on a 4x L40S cluster, sustaining a 10.5x RTF under continuous load. Trained low-latency models for English–Hindi transliteration and noise reduction for the voice-bot pipeline.

ML Engineer at Convin AI (2026-04 – Present)

• Fine-tuned a Qwen3-0.6B-based TTS model, achieving a 4.5/5 MOS in blind evaluations.
• Built an efficient inference-serving pipeline with NVIDIA Triton Inference Server on a 4x L40S cluster, sustaining a 10.5x RTF under continuous load.
• Trained low-latency models for English–Hindi transliteration and noise reduction for the voice-bot pipeline.

ML Engineer at Nurix.ai (2025-02 – 2026-04)

• Developed pipelines for SFT and DPO alignment of SLM models for domain-specific use cases, improving bot response quality.
• Built low-latency model-hosting pipelines using TensorRT-LLM, vLLM, and NVIDIA Triton Server, achieving a 30% reduction in TTFT.
• Developed an audio encoder-based user turn-end detection model using semantic and acoustic features, along with scrapers and synthetic data pipelines for audio conversational datasets.

Deep Learning Engineer (Intern) at TensorGo (2024-05 – 2025-01)

• Improved speaker diarization systems for two-speaker conversations, achieving a 12% DER score.
• Optimized model pipelines using OpenVINO for CPU inference, reducing inference time by 15% and memory usage by 40%.
• Built and deployed an automated Hindi-to-English voice dubbing application with 8% WER.

Bachelor of Technology in IT – Indian Institute of Information Technology (2020 – 2024)