Hi, I'm Aravind Durgaraj 👋

🛡️ QA Leader for AI Systems · Building Quality Gates for the Agent Era

---

🎯 The Problem I Work On

Traditional QA assumes deterministic systems. AI broke that assumption.

When the system under test is an LLM or an autonomous agent, "expected output" is a distribution, not a string. I build the frameworks, evaluation pipelines, and quality gates that make AI products shippable with confidence — combining classic QA discipline (coverage, traceability, gates) with modern AI evaluation (semantic scoring, security probing, schema contracts).

🚀 Flagship Projects

Project	What it does	Stack
🛡️ mcp-gatekeeper	The QA framework for Model Context Protocol — 46 validators × 8 gates: schema, security, functional, AI semantic eval & human review. Validates MCP servers, Claude skills & extensions before they ship.
🤖 CatalystAI	AI agent skills bundle for Zoho Catalyst — installable across Claude Code, Cursor, Gemini CLI, Copilot & Windsurf
📚 catalyst-java-sdk-wiki	LLM-optimized knowledge base — 106 docs on the Zoho Catalyst Java SDK, structured & interlinked (Karpathy LLM Wiki pattern)
🖥️ sentinel	macOS menu bar agent that keeps AI coding sessions alive while your screen stays locked

🔍 QA & Security Research in the Wild

Beyond my own projects, I actively hunt bugs in open-source AI tooling — 27 issues and 32 PRs across projects including Anthropic's claude-plugins, Google's agents-cli, Alibaba's open-code-review, PatterAI, and the W3C webmcp specification.

The work spans two sides of quality:

• 🛡️ Security research — path traversal & arbitrary file write, SSRF-guard bypasses, git option injection (RCE), pull_request_target secret exposure, fail-open policy governance, and prompt-injection-to-toll-fraud chains. Reported responsibly with CWE classification and exploit analysis; several triaged by maintainers as P0-critical.

• 🧪 Correctness & spec defects — inverted algorithm assertions and wrong struct insertions in a W3C-track spec, cosine-similarity-vs-inner-product scoring bugs, silent-failure filename mismatches, and consistency-check gaps across integrations.

This is QA where "expected output" isn't obvious — you have to reason about what should happen, then prove what does.

🧰 Tech Arsenal

Quality Engineering

Languages & Runtime

AI Ecosystem

📊 GitHub Stats

💼 What I Bring as a QA Leader

class AravindDurgaraj(QALeader):
    def strengths(self):
        return [
            "Quality processes for non-deterministic systems",
            "QA tooling developers actually adopt",
            "46-validator pipelines > manual checklists",
            "Bridging classic QA rigor with AI evaluation",
        ]

    def philosophy(self):
        return "If you can't gate it, you can't ship it."

---

⚡ Open to QA Manager / Lead AI QA Engineer roles

Let's talk about making your AI products testable.

<img

src="https://raw.githubusercontent.com/Sushmitadasari/Sushmitadasari/main/dist/github-jet.svg"

alt="GitHub Jet Heatmap"

width="100%"

/>