At Listen2It, I fine-tune and adapt generative models for neural avatars, speech, and video. On GaussianTalker, I integrated audio-visual sync supervision and audio-to-motion conditioning, raising lip-sync accuracy to an 8.5 Sync-C score and removing head jitter.
I also bring research models into production: I reduced Qwen3-TTS inference to about 0.2 RTF on a nano-vLLM serving stack and shipped a RunPod Serverless vocal-separation pipeline merged with ASR and Seed-VC endpoints. At ZettaZeal Technologies Pvt. Ltd., I fine-tuned diffusion and GAN models for virtual avatars and try-on, and developed computer-vision applications.
