AI Security Architect | Senior Cybersecurity Engineer | Agentic AI Security
CISSP, CISM and CRISC certified cybersecurity professional working where enterprise security meets AI systems and automation.
I build AI agents for security operations and the controls that keep them safe to run: triage and investigation agents, tool authorization with human approval, retrieval with citations, and evaluation harnesses that measure what each piece actually does.
- Agents for SOC triage, investigation and incident response, with people kept on every decision that carries weight
- Tool authorization for agents: MCP gateways, privilege tiers and human approval
- Evaluation harnesses with labeled data, clean-versus-injected twins and predictions committed before runs
- Detection engineering with Microsoft Sentinel and KQL
- AI security controls on Microsoft 365, Azure and Google Cloud
LLM-driven SOC alert triage using ASIM-normalized alerts, a labeled evaluation corpus and a measured test harness.
Measured result: 90% accuracy with a 0% false-close rate on the project evaluation set, reproduced at 90.2% over 500 triages in the Nemotron lab.
Nevada cut the casino cyber incident notification window from 72 hours to 24 in January 2026. This assistant investigates incident telemetry, builds an evidence timeline and drafts the regulator notice with citations, behind the IT Operations MCP Gateway. A named person decides whether to notify and approves every send; the agent can do neither.
First model runs: Nemotron 3.5 Lightning reached 5 of 6 conclusions but missed in the dangerous direction, reporting the attacker gone when that was not yet known. Claims of absence now need positive evidence; on the rerun its one miss was over-cautious instead, and a refused claim sent it to examine the gaming systems first. Both runs and their limits are in the build log.
Evaluation harness for NVIDIA Nemotron 3.5 Lightning, qwen2.5 and Gemini 3.8 Flash on SOC alert triage and IT-operations agent tasks. Measures accuracy, prompt-injection effects against clean twins, and whether a trusted MCP authorization gateway holds when a model is steered.
Measured result: across 216 agent episodes, 96 of them carrying injected instructions, the payloads steered three of four models into requesting privileged actions and the gateway allowed none of them. Predictions for the key experiments were committed to git before any run.
Natural-language-to-KQL benchmark for security questions.
Measured result: schema grounding raised execution success from 9.1% to 68.2% on a local 14B model, measured against a real Kusto engine on the project test set.
Secure MCP gateway for AI agents with four-tier authorization, human approval for privileged actions, prompt-injection defenses, append-only auditing, and ITSM and identity integration patterns.
Measured result: no privileged or denied action executed in 216 agent episodes run against it by the Nemotron lab, including every episode in which the model was steered. Release v1.1.0 requires issued credentials on the HTTP side, so approvals need a key agents are never given.
A declarative Microsoft 365 Copilot agent grounded in SharePoint, calling a controlled Security Actions API with Entra token validation, READ / WRITE / PRIVILEGED / DENIED actions, independent human approval and a tamper-evident audit chain.
Measured result: 6 of 6 deterministic control-plane cases and 7 of 7 live Microsoft 365 security checks passed. Indirect prompt injection, a document posing as an approval, and agent self-approval were all blocked.
A RAG assistant on Vertex AI and Cloud Run, attacked and then secured one control layer at a time across nine modules: IAM, Model Armor, VPC Service Controls, Sensitive Data Protection, Cloud Armor, CMEK, detections and identity-scoped retrieval.
Measured result: with no controls, 17 of 20 attack prompts and 8 of 10 harmless questions leaked synthetic SSNs and card numbers through a poisoned document. Model Armor alone stopped the leaks but blocked all 30 requests. Adding a DLP ingestion pipeline kept leaks at zero and let harmless questions through again. One finding: a Model Armor setting reported "Enabled" while only inspecting traffic, not blocking it.
Local clinical RAG over a medical reference manual, with hybrid retrieval, domain cross-encoder confidence, two-layer abstention, and an evaluation harness scored by two independent judges.
Measured result: on 48 golden items, the abstention gate refused all 6 out-of-corpus questions while serving 98% of answerable ones, with 90% citation coverage and no invalid citations. A validation run that seemed to show both judges missing all 10 injected hallucinations turned out to be scoring stale verdicts; judged correctly, MiniCheck caught 10 of 10.
- AI Security Platform: a Rust prompt-injection detection engine and a Go security gateway sharing one detection core. A CI gate runs a labeled corpus on every commit: 25 of 26 attacks detected and 0 of 15 benign inputs flagged; the miss is character-spaced text.
- Remediation Recommender: two-stage recommendation engine with collaborative filtering and LightGBM LambdaRank, a SQL feature pipeline, an evaluation gate and FastAPI serving.
- AI Agent Automation: multi-stage LLM workflow that generates, reviews and revises structured content.
Agent security
Identity → Authorization → Tool permission → Human approval → Execution → Audit
RAG security
User → Retrieval → Untrusted content → LLM → Policy enforcement → Response
Adversarial testing
Prompt injection · Indirect prompt injection · Tool abuse · Privilege escalation · Data exfiltration · Unsafe agent actions
AI governance
NIST AI RMF · MITRE ATLAS · OWASP LLM Top 10 · Secure by design · Zero Trust · AI TRiSM
Enterprise security architecture and engineering across EDR/XDR, endpoint security, vulnerability management, IAM, SIEM, security automation, cloud security, compliance and incident response.
Platforms and tools: Microsoft Azure · Google Cloud · Microsoft Sentinel · Microsoft Defender · Tanium · CrowdStrike Falcon · SentinelOne · ServiceNow · Splunk · Tenable · Python · PowerShell
These projects trace the path from applied machine learning into AI security engineering.
- Predict House Prices: linear regression
- Real-Time Fraud Detection: Apache Spark
- Loan Default Prediction: XGBoost
- Logistic Regression from Scratch: SGD