Ashray Saini

I'm building high-throughput AI infrastructure at SKETS Studio, where I engineered vLLM inference on NVIDIA B200 GPUs, deployed quantized Llama, DeepSeek, and Gemma models, and benchmarked concurrent generation using TTFT and ITL metrics.

Previously at TrueFan AI, I trained VQ-VAE and StyleGAN models for facial and teeth reconstruction, improved viseme synthesis with LoRA and LoHa adapters, and built Stable Diffusion pipelines for targeted inpainting and image-editing APIs.

At Dresma and i2V Systems, I delivered agentic RAG, ETL, REST API, video analytics, traffic intelligence, face recognition, segmentation, and edge inference systems. My work spans TensorRT, OpenVINO, ONNX Runtime, AWS, MongoDB, FastAPI, and production computer vision pipelines.

I'm particularly interested in generative vision, multi-agent reasoning, asynchronous data systems, and making deep learning models faster and more reliable in production.