Bengaluru · Applied AI Research & Engineering
Arun Patil Labs builds and advises on LLM inference — speculative decoding, hardware-aware optimization and model bring-up on non-NVIDIA silicon — and turns it into products that serve real users at low latency and cost.
Start a conversationSpeculative decoding, quantization and serving design that cut latency and cost per token without hurting quality. Model bring-up across architectures such as Qwen3, GLM-4 and GPT-OSS.
Routing between small, fast models and frontier models for voice agents, built for Indian languages and real-time call latency budgets.
Research into block diffusion, reasoning-efficient transformers and data-efficient training, taken from experiment to deployable model.
Production vision systems for retail and operations: inventory, loss prevention, safety and compliance at multi-site scale.
Founded by Arun Kumar Patil, whose work spans frontier LLM inference at Cerebras, FPGA AI kernels, CNN compilers and quantization at Xilinx, a production computer-vision platform at Target, and embedded imaging at Ittiam Systems.
Published research includes the Implicit Chain Transformer (“Thinking Inside the Box”).