Sanket Deshpande

Background

About Sanket Deshpande

I build and ship AI products end-to-end — from LLMs and multimodal systems to distributed training and real-time inference. My edge: speed, ownership, and turning research into production. Shipped AI Products Cre8able — Multimodal video-editing platform (beta in 6 weeks). Cut creator editing workload by 70%, delivering <2s clip latency with Gemini + FFmpeg + FastAPI/S3. Whisker — LLM-powered pet-care assistant (shipped in 2 weeks). Built a containerized RAG backend (Postgres + FastAPI) with sub-second response streaming; Applied ML Systems Built 3D CNN–Transformer hybrids for cognition prediction on 12k+ MRI scans, boosting performance +15% vs. baseline CNNs. Optimized multi-GPU distributed training (4× V100, mixed precision), achieving 7× faster epochs. Stack and Skills ML/AI: PyTorch, Transformers, CLIP, multimodal pipelines LLMs: Hugging Face, LangChain, Pinecone/FAISS, custom RAG pipelines Production: FastAPI, Docker, Postgres, async/streaming APIs, Git Scaling & Infra: DDP, SLURM, AWS/GCP, multi-GPU training, real-time inference pipelines I thrive in early-stage, high-velocity environments where ownership, speed, and impact matter. If you’re building the next AI product and need someone who can turn an idea into a shipped system in weeks, let’s connect.