← All links

Ar. Yash Arora

AI Systems Engineer · LLM Orchestration, Systems Integration & Audit

I build AI systems that run on their own — and then I check whether they actually work together.

My work has moved through three stages: building individual tools, connecting them into systems that close a whole problem loop, and now auditing those systems so the parts genuinely reinforce each other instead of merely coexisting. That middle layer — orchestration, failure handling, cost control — is the part most teams under-invest in, and it is the part I own.

I trained as an architect, which is where the instinct comes from: design the constraints first, then let the structure hold.

What I build

Agentic systems & orchestration
Multi-agent architectures where agents hand off work, share state and recover from failure — separating LLM reasoning from deterministic, tested execution.
Multi-provider LLM routing
Provider abstractions with automatic failover that classify errors as retryable or terminal, so an outage degrades service instead of stopping it.
Cost engineering
Fail-closed budget controls, capability-tier routing and content-addressed caching. One pipeline’s worst-case cost went from ₹239 to ₹6.34; a routing audit caught an 11.5× model overspend.
Governance in code
Data-classification gates that block sensitive payloads before they reach a third-party API, and spend gates that demand written justification. A documented rule nothing enforces is only a preference.
Production web & automation
Live Next.js and Astro applications, hybrid local/cloud schedulers on GitHub Actions, and publishing pipelines across six platforms with idempotent dispatch.

Selected work

Live · Astro · Tailwind · GSAP · Cloudflare Pages
Production site for my Architecture × AI studio, shipping minimal JavaScript. Front-end iterated in Google AI Studio with a diff-based reconciliation step before anything is integrated.
LLM Routing & Cost Governance
18-model catalogue · 4-class data policy · enforced in code
A routing law deciding which model may process which data class. Providers were disqualified on evidence — one free tier banned because its terms permit human review of API inputs.
The Researcher — Credibility Instrument
Multi-stage vision + LLM pipeline
Maps who asserts a claim, who opposes it, and each source’s independence — built to assemble evidence and deliberately issue no verdict. Scope was cut after a pilot caught it treating an accused party’s denial as credible evidence.
Explanation-Video Production System
Chunked render · 8m47s · quality-gated · ₹0 external spend
Programmatic video pipeline with subprocess isolation and FFmpeg assembly. Every beat passes an 8/8 quality gate, including audio-sync verified to a 0.0005-second delta.
Live · Next.js · Vercel · Cloudflare DNS
Built on a permanence contract: the printed QR encodes only the domain root, and every destination resolves through an indirection layer — so links can change forever without reprinting anything.

Background

FounderShakti Yukti — Architecture × AI studio (2026 – present)
Co-FounderShakti Kreeda (2025 – concluded Feb 2026)
EducationBachelor of Architecture, AIT-SAP, Greater Noida (2020 – 2025)
CredentialLicensed Architect (India)
EngineeringSoftware, AI and systems — self-taught, shipped to production
Download full background (PDF)

Last updated 2026-08-04 · about.aryasharora.vip