Open to opportunities

Mehrin Kiani

@mehrinkiani

Principal machine learning engineer building secure-by-design LLM systems with production-grade reliability and measured safety.

Canada

Message

What I'm looking for

I’m looking for a team where I can architect secure-by-design LLM/agent systems, combine deterministic policy enforcement with rigorous evaluation, and lead high-impact modeling and mentorship that translate research into reliable production safety.

I architect and lead “secure-by-design” AI systems, bridging SOTA research with production-grade reliability for high-stakes deployments. I focus on OOD robustness, multilingual grounding, and building deterministic safety systems that still let models remain flexible.

At Palo Alto Networks, I spearheaded an in-house agentic security framework that decouples data and control planes for enterprise safety with deterministic enforcement. I delivered automated security enforcement achieving 93–100% attack detection with sub-2ms latency across 16 security scenarios, and built profiling systems that reverse-engineer agent capabilities without source code.

I also designed policy-generation and runtime enforcement pipelines—turning unstructured knowledge into machine-enforceable Rego policies and combining deterministic Rego evaluation with semantic LLM analysis for high-risk operations. My evaluation framework tested 4 production agents across 100+ OWASP-mapped scenarios, and I led multilingual synthetic data strategies to close performance gaps in low-resource languages.

Earlier, as a founding ML scientist at Protect AI (acquired by Palo Alto Networks), I reported to the CTO and built the core scanning engines that secured thousands of models and drove early customer acquisition. I also led LLM security research, including gradient-based adversarial attacks and LoRA fine-tuning for prompt-injection and jailbreak detection, while mentoring teams toward stronger, embedding-based robustness.

Experience

Work history, roles, and key accomplishments

Current

Principal Machine Learning Engineer

Current

Palo Alto Networks

Jul 2025 - Present (11 months)

Architecting secure-by-design LLM deployments with an agentic security framework that decouples data and control planes for deterministic enterprise safety. Delivered 93–100% attack detection across 16 scenarios with sub-2ms latency and achieved 98% F1 on multiple public benchmarks via fine-tuning of Gemma and Phi models.

LLM Security Out Of Distribution (OOD) Robustness Synthetic Data Generation Multilingual Grounding

Founding ML Scientist

Protect AI

Aug 2022 - Jul 2025 (2 years 11 months)

Reported to the CTO to pioneer Security-by-Design and built core AI model scanning engines to power security enforcement and early enterprise lead generation. Identified 3,300+ vulnerable models with Guardian, advanced ModelScan for Keras/PyTorch/TensorFlow, and improved jailbreak detection by 13% through LoRA fine-tuning for prompt-injection defense.

Model Vulnerability Scanning Hugging Face Integration Keras BPE Tokenization

ML Researcher

University of Essex

Jan 2019 - May 2022 (3 years 4 months)

Conducted research on explainable machine learning methods, including time-dependent explainable AI, multivariate pattern analysis, and regularized cognitive maps for neuroimaging signal analysis. Published work in IEEE Transactions on Artificial Intelligence, Nature Communications Biology, and IEEE Transactions on Cognitive and Developmental Systems.

Explainable AI Time Dependent XAI (TXAI)Multivariate Pattern Analysis (MVPA)Cognitive Maps Type 2 Fuzzy Systems FNIRS Signal Analysis

Education

Degrees, certifications, and relevant coursework

University of Essex

PhD in Computer Science, Computer Science

2019 - 2022

Activities and societies: Doctoral scholarship; research in TXAI, MVPA, and cognitive maps; published work in IEEE and Nature-related journals.

PhD in Computer Science focused on developing novel explainable machine learning algorithms, including time-dependent explainable AI (TXAI), multivariate pattern analysis (MVPA), and cognitive maps. Research resulted in publications in venues such as IEEE and Nature Communications Biology.