Oliver Grainge

Experience

• Contract Researcher — Performance Engineering Feb 2025 – Present
  Arm · Remote
  Architected 6 hands-on tutorials for the Arm Total Performance toolkit covering memory optimization, library acceleration (APL, KleidiAI), and automated porting for AWS Graviton instances.
• Research Assistant Jun 2025 – Nov 2025
  University College London · London, UK
  Engineered 1.58-bit precision pipeline for Stable Diffusion (4x memory reduction, 95% quality retention). Designed custom CUDA and Triton kernels for bit-packed tensor operations, delivering 30% speedup over PyTorch baseline.
• Visiting Researcher Aug 2024 – Jan 2025
  Queensland University of Technology · Brisbane, Australia
  Engineered speculative decoding for vision-language transformers achieving 2.5x inference speedup for sub-100ms robotic navigation. Implemented training data filtering methods achieving equivalent accuracy with 38% less data.
• Research Fellow Jan 2024 – Jan 2025
  AI Security Institute · Remote
  Built automated benchmarking framework evaluating 25+ VLMs across 26k geo-tagged images with 99.9% reliability over 500k+ API calls. Developed privacy-preserving techniques reducing geolocation accuracy by 40%, with interactive demo attracting 5k+ users.
• Contract Researcher — AI Inference Optimization Nov 2024 – Jan 2025
  Arm · Remote
  Engineered demonstrations achieving 40% latency reduction via SIMD/INT8 on mobile and 2.1x throughput on cloud instances. Built Hyperopt-based per-layer precision optimizer demonstrating 22% memory reduction on GPT models.

Selected Publications

• Assessing the Geolocation Capabilities, Limitations and Societal Risks of Generative Vision-Language Models
  AAAI 2025 · First comprehensive benchmark of VLM geolocation capabilities across 4 datasets
• TeTRA-VPR: A Ternary Transformer for Compact Visual Place Recognition
  IEEE RAL · 65% memory reduction and 35% latency reduction for VPR transformers
• Design Space Exploration of Low-Bit Quantized Visual Place Recognition
  IEEE RAL · Deployment guidelines for extreme quantization on embedded devices
• Structured Pruning for Efficient Visual Place Recognition
  IEEE RAL · Channel pruning achieving 21% latency and 16% memory reduction with <1% accuracy loss

Open Source Projects

• BitOps
  C++ · CUDA · ARM NEON · AVX2
  High-performance ternary matrix multiplication library with multi-backend support. 16x memory reduction via 2-bit weight packing with kernels outperforming PyTorch FP32 on edge devices.
• BitCore
  PyTorch
  Quantization-aware training toolkit with drop-in BitLinear layers (BitNet, TWN, ParetoQ 1.58-bit). Train-to-deploy workflow with optional BitOps acceleration for 8x inference memory savings.
• BitNet Chat Interface
  Python · Gradio · CUDA
  Interactive chat with 1.58-bit BitNet models. 24x speedup and 80% memory reduction on ARM M4 vs PyTorch FP32, with backend switching and streaming responses.
• VSLAM
  Python · OpenCV · NumPy
  Pure-Python stereo SLAM pipeline (KITTI-compatible) with feature tracking, stereo matching, PnP/ICP motion estimation, and bundle adjustment.

Technical Skills

Education