Dhairyashil R. G.
I find out where the time, the cost and the failures actually go, in compilers, parallel code and AI agents, and I publish the measurement so someone else can re-run it.
Senior Member of Technical Staff, AMD · uProf profiler. PhD, IISc Bangalore.
Compilers
Seven upstream LLVM/MLIR pull requests, four merged, on vector bounds reasoning and NVVM lowering. One of them removes every masked load from a PolyBench kernel on AArch64, 14 → 0sourcehttps://github.com/llvm/llvm-project/pull/215340checked2026-09-22hardwareAArch64 NEON, Apple silicon (no SVE); not measured on SVEcommandbench/run.shkindmeasured, and the kernel runs about 5×sourcehttps://github.com/llvm/llvm-project/pull/215340checked2026-09-22hardwareAArch64 NEON, Apple silicon (no SVE)commandbench/run.shkindmeasured faster. Reproduce it →
Parallel code
Tech lead on a C++/MPI particle simulator that ran 2×sourceSankhyaSutra Labs internal benchmark (2018–2024), résuméchecked2026-09-21kindself-reported faster than Sandia's SPARTA on the same hardware, scaled past 100sourcerésumé, SankhyaSutra Labschecked2026-09-21kindself-reported nodes.
AI agents
A video series and four studies with public run ledgers. In one, the same public prompt injection succeeded 2 to 35 of 97sourcenext_series/T02_agent_security/series/FACTS.md in dhairyashilRG/agent-harnesses-2026checked2026-09-19kindmeasured times, depending only on the model. Watch → Data →
Latest from the log
Episode 3 is on YouTube: the anatomy of a harness, the seven places bugs hide.
Episode 2 is on YouTube: the loop every agent runs, think, act, observe.
Episode 1 of the harness series is on YouTube: what an AI agent harness is, and the same model doing better with a different one.
Everything here is meant to be run by someone who is not me.