Khush Patel

02 Experience

03 Selected Work

• EthicalEye, a cross-lingual AI ethics model. Open-source content moderation model supporting 88 languages, built on XLM-RoBERTa. 10,000+ downloads on Hugging Face; adopted by companies for production content moderation, reducing manual review by approximately 70%. Released under Apache 2.0.
• AgentDefender, a neural-embedding defense against prompt injection. Novel methodology and benchmark for securing LLM-based agents, deployed on Lyzr's production platform and published (see section 07).
• Regulatory knowledge graph for financial compliance. Knowledge graph spanning FINRA, FRB, OCC, EU AI Act, DORA, DFS, FDIC, and NAIC regulations, powering a continuous compliance monitoring system that replaced sampling-based audits.
• Lyzr Agent Studio and Skott. No-code enterprise AI agent platform (522 Product Hunt upvotes) and an autonomous AI marketing agent producing 1,000+ pieces of content with no human intervention (424 Product Hunt upvotes).
• Responsible AI engine for the Lyzr platform. Built Lyzr's production Responsible AI module: PII detection and redaction, toxicity metering, NSFW, gibberish, and prompt-injection detection, topic controls, and pluggable policy engines (OPA, Cedar, AWS Bedrock, and Google Vertex guardrails). Also integrated it into ComputerAgent's open-source and enterprise editions as a chained guardrail redaction proxy with policy-bound agent runs.
• Agent Simulation Engine and Agent Improvement Engine. Created at Lyzr: a CI/CD-style platform for AI agents that tests agents against realistic simulated scenarios, evaluates results with AI-powered judgment, analyzes failure patterns, and automatically improves agent configurations using reinforcement learning techniques.
• ShadowLM Trainer. Creator and maintainer of an open-source fine-tuning SDK for open models: 13 training methods, any hardware, batteries-included install. Built at Lyzr Research Labs. GitHub.
• OpenGAP (the Git Agent Protocol), GitAgent, and ComputerAgent. Contributor to the open-gitagent ecosystem: OpenGAP, a framework-agnostic, git-native standard for defining AI agents with first-class FINRA, Federal Reserve, and SEC compliance support (2,800+ GitHub stars); GitAgent, a git-native AI agent whose identity, memory, and skills are version-controlled files (590+ stars); and ComputerAgent, a portable agent runtime and reference implementation of the Harness Protocol. OpenGAP · GitAgent · ComputerAgent.

04 Community Recognition

05 Awards & Recognition

06 Advisory Work

07 Research & Invited Talks

Full record on Google Scholar.

• "The Six Sigma Agent: Achieving Enterprise-Grade Reliability in LLM Systems Through Consensus-Driven Decomposed Execution" (2026). arXiv:2601.22290.
• "Cognis: Context-Aware Memory for Conversational AI Agents" (2026). arXiv:2604.19771.
• "LogNEO: A GPT-Neo Reinforcement Learning Framework for Accurate Real-Time Log Anomaly Detection" (2026). arXiv:2606.08153.
• "CAMP: Compositional Agent Modeling Protocol" (2025), preprint.
• "AgentDefender by Lyzr: A Benchmark Evaluation and Neural Embedding Approach for Agent Prompt Injection" (2024). Available on ResearchGate.
• "ReCoR-X: Reflective Comparative Reward for Agent Reinforcement with Experts" - in preparation; reinforcement learning framework for expert-guided agent training.
• Code generation and documentation using large language models (2023-2024), research conducted at the Advanced Engineering School.

08 Education