About
Background, architectural philosophy, and engineering toolkit.
I am Sudhanwa Kaveeshwar, an engineer focused on the intersection of deep learning inference systems, distributed data infrastructure, and generative AI architecture.
Over the past several years, I have worked on scaling high-throughput streaming systems, writing custom CUDA kernels for attention acceleration, and designing evaluation pipelines for multi-agent LLM systems.
My engineering work is guided by the principle of mechanical sympathy—understanding how hardware, memory hierarchies, and operating system caches function so we can write software that fully utilizes compute resources rather than masking inefficiencies with excessive cloud infrastructure.
Core Engineering Principles
Profile instruction stalls, memory bandwidth saturation, and cache misses before making architectural assumptions.
Prefer single-process vectorized engines over sprawling distributed clusters when local memory bandwidth suffices.
Avoid vibe-driven LLM development. Enforce unit-tested rubrics, statistical significance, and ground-truth assertions.
Eliminate serialization overheads at ingestion boundaries using Apache Arrow and cache-aligned record batches.