about

I build from first principles. Codecs, compilers, inference engines, audio software, game engines. I would rather understand a system all the way down and rebuild it than wire libraries together. Most of my work lives where audio, machine learning and systems programming meet.

• Neural speech codec — learns a compact latent representation of speech and reconstructs it with a neural vocoder, instead of hand-designed signal coding.
• QLAC — a learned audio codec on quantised latent representations (vector quantisation plus residual coding).
• C++ neural inference engine — real-time on a single core, faster and lighter than TFLite and ONNX, fully on-device.
• Flow — my own statically typed compiled language with backends for LLVM/MLIR, SPIR-V (Vulkan/CUDA), Metal and WebAssembly.
• Umbra — a 2D procedural rendering engine with GPU acceleration. RayTherion — a real-time GPU ray tracer.
• Zaza — a parallel build system in Zig, a modern alternative to CMake.
• Autonomous drone systems — city-scale C++ simulation and control with realistic flight dynamics.

I run Quilio, my audio technology company, shipping real-time C++ audio to hundreds of thousands of systems. Partitioned convolution, low-latency neural vocal processing, Opus streaming, spatial audio. Core engine work on Tracktion alongside the creators of JUCE, plus a set of open-source JUCE plugins.

Real-time decoding of imagined speech and intent from neural signals, on implant and embedded hardware. Plus closed-loop neuromodulation: brain-responsive music that drives targeted, non-invasive neurostimulation.

Around 80 independent research preprints across machine learning, audio and dynamical systems at research.abhishek-shivakumar.com. All preprint or in progress, not yet peer-reviewed.