Asma Beevi is a senior engineer at NVIDIA developing NVIDIA TensorRT Model Optimizer, a toolkit for making powerful AI models smaller, faster, and less expensive to run. Her work concentrates on preserving model accuracy while reducing the precision, memory, and computation required for deployment.
Also publishing as Asma Beevi K. T. and Asma Kuriparambil Thekkumpate, she coauthored 2022 research on millimeter-wave radar hand-shape recognition using deformable transformers and three-dimensional point clouds. She subsequently worked on pruning computer-vision models before turning to quantization as large language models reshaped the computational demands of AI.
At NVIDIA, she has contributed to Model Optimizer, which supports quantization, pruning, distillation, and other deployment optimizations. Her coauthored technical work spans low-precision inference, diffusion models, memory-efficient fine-tuning, and accuracy recovery for compressed language models.
Architecture-sensitive quantization: Beevi treats compression as a way to deliver greater intelligence within a fixed computing budget. She works with NVFP4, a four-bit floating-point format that shares scaling information across small groups of values, and emphasizes mixed-precision strategies: mixture-of-experts components can often tolerate aggressive compression, while attention projections may require higher precision. Benchmarking remains essential because useful compression must preserve actual model performance.
Quantization-aware training: When post-training quantization reduces accuracy, Beevi’s work on training-based recovery describes fine-tuning models under simulated low-precision conditions, sometimes with guidance from a higher-precision teacher. She also coauthored a four-bit gpt-oss adaptation workflow that restores higher precision for supervised fine-tuning before returning the model to efficient deployment.
BeeLoop and persistent agent systems: Outside model compression, she created BeeLoop, an open-source personal assistant and loop-engineering project, and maintains AgentInProduction. Her writing on agent-loop architecture explores memory and event-driven behavior beyond one-shot coding assistance.
At the 2026 AI Engineer World’s Fair, Beevi outlined a broader optimization agenda encompassing KV-cache compression, long-context reasoning, activation sparsity, and the challenges introduced by increasingly heterogeneous model architectures.
Making a model fit is only the beginning: selective precision, architecture-aware evaluation and runtime compression determine whether local intelligence remains useful.
