Based in San Francisco, California. Available nationwide, open to relocation.
I write the controls, then I write the code that tests them.
At KPMG I spent two years on IT audit, third-party risk, cloud and IAM review, and SIEM and DLP validation across seven regulated financial institutions. Since then I have been building the automation for that work: an LLM audit agent at Avaly.AI running 227 controls, a CI/CD pipeline that blocks vulnerable code at merge, and a 30-control framework for LLM systems out of my thesis.
GRC and IT audit
Third-party risk
AI security
DevSecOps
Detection and response
Selected work
Six things I would want to walk you through.
Problem first, then what I built, then what it changed. If a claim has a number on it, I can show you the workpaper, the repo or the paper behind it.
Avaly.AIJun to Aug 2025AI Security Engineer
A compliance team that could not keep up with the product it was governing.
The problemAn early stage AI company needed to show governance across hundreds of controls, with evidence collection done by hand. Manual assessment does not scale next to a product shipping weekly, and the gap between what the policy claimed and what the system did was widening.
What I builtAn LLM Audit Agent in Python and FastAPI, containerized with Docker and deployed through GitHub Actions CI/CD, automating evidence collection, control validation and compliance documentation across 227 controls mapped to NIST AI RMF and ISO/IEC 42001.
The security work alongside itStructured OWASP LLM Top 10 threat modeling across five core product workflow stages, plus two internal tabletop exercises simulating model misuse, unreliable output and sensitive data exposure.
- 10+ exploitable LLM attack vectors surfaced and 12+ safeguards shipped, including prompt injection, evidence tampering and sensitive data exposure, each mapped to a NIST AI RMF or ISO/IEC 42001 control
- Tabletops produced defined escalation paths and response ownership for high-risk AI scenarios, which the team did not have before
- GenAI vendor assessment standardized across 7 trustworthiness domains, adopted as the company process for new vendor onboarding
KPMG IndiaJan 2023 to Jul 2024Cybersecurity Analyst7 regulated institutions
Audit at scale, where every finding has a regulator attached to it.
The workRisk-based technology audit across RBI and SEBI regulated financial institutions: scoping, control walkthroughs, test plan development, fieldwork, issue validation and remediation follow-up spanning cloud, IAM, network, Windows, datacenter and SDLC environments.
Where the value landedThe deliverables outlived the engagements. Remediation dashboards and risk registers became each client's primary mechanism for tracking observation closure, and five policies I authored were accepted at CISO level and referenced in later supervisory reviews.
- Control maturity elevated from 2.5 to 3.8 ahead of supervisory review, driving 80+ identified gaps to remediation against NIST CSF, ISO 27001 and RBI mandates
- 35+ audit workpapers prepared and reviewed, documenting testing procedures, evidence and conclusions
- 5 enterprise policies accepted at CISO level across Asset Management, Change Management, BCP/DR, Security Awareness and Tabletop Exercises
- 10+ identity and access governance gaps surfaced across 5 Active Directory environments
- 20+ SDLC and change management control tests across 4 engagements, producing 8+ formal observations for management action
- Cloud security assessed across AWS, Azure and GCP at 4 institutions, plus 8+ network gaps found across 3 institutions in firewall, NAC and segmentation controls
- RSA Archer GRC deployed at 2 institutions, configuring assessment templates and remediation workflows that became the auditable reporting record
Third-party riskKPMG7 engagements
Vendor assessments owned from scoping through remediation closure.
The workThird-party and vendor security assessments across regulated institutions. I led pre-assessment preparation and scoping, collected and reviewed vendor documentation and playbooks, identified gaps, then took remediation discussions directly to service providers rather than handing a report over the fence.
What made it workRelationship management, honestly. Assessment readiness depends on vendor managers and service providers actually answering, so I kept the communication line open from kickoff through closure and handled the information security requirement questions myself.
- Assessed controls spanning infrastructure, access management, physical security, application security, change management and business continuity
- Datacenter walkthroughs covering physical access governance and Windows event-log monitoring
- Risk observations escalated to client security teams and KPMG leadership, then tracked through closure
M.S. thesis and research2025 to 2026Under review, ICSE 2027
An auditor can read NIST AI RMF cover to cover and still not know what to test.
The problemNIST AI RMF and ISO/IEC 42001 say what outcomes to achieve. Neither says what evidence proves it or what a failing control looks like. Teams shipping LLM features get asked to demonstrate governance with nothing concrete to demonstrate against.
What I builtA 30-control framework cross-mapped across NIST AI RMF, ISO/IEC 42001 and OWASP LLM Top 10 (2025), anchored to a 0 to 5 maturity model, then a browser-based audit agent that runs it.
- 81.1% agent-to-human agreement against a 72.2% inter-assessor baseline, which is the more interesting number: assessor disagreement is the real bottleneck in audit
- Research paper under peer review at ICSE 2027, with a second submission in preparation for IEEE CSF 2027
SecurePipeMar 2026DevSecOps
Tooling that reports findings changes nothing. Tooling that blocks merges changes behavior.
What I builtA GitHub Actions pipeline across a Python Flask application with Bandit for SAST, Semgrep, pip-audit for SCA and OWASP ZAP for DAST, running on every commit. Branch protection requires all five pipeline jobs to pass before anything merges to main.
- 16 CVEs across 4 packages including denial of service, path traversal and sandbox escape, all eliminated through targeted version remediation
- 2 high-severity findings reduced to zero: command injection (CWE-78), debug mode remote code execution (CWE-94), SQL injection (CWE-89) and a hardcoded secret (CWE-259)
SOC analyst capstone2026Detection engineering
You cannot judge detection coverage without generating the telemetry yourself.
What I builtA phishing-to-credential-dump attack chain simulated end to end in an isolated VirtualBox lab, Windows 10 victim against a Kali attacker running Metasploit, with Sysmon and Winlogbeat producing telemetry across the kill chain. Then I wrote the detections for it.
- Detection rules authored in Elastic SIEM and Sigma, mapped to MITRE ATT&CK T1566, T1059 and T1003, with alert fidelity validated against generated activity
- Full analyst workflow run end to end: log ingestion, alert triage, forensic pivoting, IOC identification, containment documentation and formal incident reporting
Coverage
Pick the track you are hiring for.
Each one shows how I would describe the work, where it shows up, and what I am hands on with versus what I only have working knowledge of.
Certifications
CompTIA Security+ISO/IEC 27001 AssociateCISA, exam scheduled September 2026AWS Certified AI Practitioner, in progressMulti-Cloud Red Team Analyst, Cyberwarfare LabsEC-Council Network Defense EssentialsEC-Council Ethical Hacking EssentialsCTF winner, UWB GreyHats
Foundational, not hands on
FedRAMP, DPIA, NYDFS 23 NYCRR 500, ISO 27701 and NIS2. I know the shape of each and would get productive quickly, but I have not owned one end to end. I would rather tell you that now than in week three.