Arsh Goyal businessarshgoyal

Hi, I'm Arsh Goyal 👋

Senior Software Engineer · Open-source contributor · I make LLM systems faster, safer, and more reliable.

Applied AI infrastructure — inference tuning, agent frameworks, RAG pipelines, evals.
Upstream open source — I fix real bugs in the tools I use and send the patch back.
Systems that stay up — SRE tooling, workflow orchestration, observability.

I like problems where a measurement settles the argument: a benchmark, a repro, a failing test turned green.

Project	What it is	Stack
armtune	Auto-tunes `llama.cpp` for Arm64 and proves it on GitHub's free Arm runners — 5.03× faster decode, 66% smaller model, fully reproducible from the Actions tab.	Python · C++ · KleidiAI · GGUF
scopecheck	Eval harness measuring whether tool-using agents respect a declared permission boundary when crossing it would be easier. Ships with a 5-page writeup.	Python · LLM evals · AI safety
sre-coworker	Alert → citation-backed root-cause brief → human approval → Jira ticket → Devin drafts the fix PR. Deterministic triage, no LLM in the loop until the fix.	Python · Pub/Sub / SQS / Redis · Jira · Devin API
opssentinel	Autonomous operations analyst on Snowflake: dynamic-table anomaly detection, Cortex Search for root cause, human-approved actions. Snowflake CoCo CLI Hackathon 2026.	Snowflake · Cortex · Python
ratelab	Interactive playground to visualize and compare rate-limiting algorithms in real time, with copy-paste implementations.	TypeScript

Patches I've landed or have in review across the AI / infra ecosystem:

Merged

PHPOffice/PHPPresentation — Apple Keynote (.key) reader and writer
docling-project/docling — keep WebVTT cue text ordered after multi-line spans
deepset-ai/haystack — standard split metadata for DocumentSplitter(split_by="function")
thomaspinder/GPJax — consolidate duplicated paramax-unwrap helper · #684 remove unreachable compat shim

In review

📍 India · Currently open to senior / staff roles in AI infrastructure and platform engineering.