About
Production AI for cloud security — founded, architected, and shipped.
I'm a Software Engineering Manager and Principal ML Engineer at Fortinet (FortiCNAPP), leading a team across detection, generative AI, and data platform. I founded the product's AI Assist line, helped found and architect the engine that correlates hundreds of detections into single high-fidelity incidents, co-designed small-language-model detectors (transfer learning; supervised + unsupervised anomaly detection) that feed it, and led its migration onto an open lakehouse. I care about systems that are not just smart but trustworthy and cheap to run at scale — and about turning model performance into measurable business outcomes.
Before this I was a Senior Data Scientist at Salesforce and a Senior Staff Data Scientist at HPE. I hold an MS in Analytics from Arizona State (GPA 4.0, Beta Gamma Sigma), a granted patent, an allowed patent on generative-AI event processing, and a patent pending.
Selected Work
What I've built
A few of the things I've led at Fortinet / Lacework — plus the decade of production ML before it. Numbers are real and measured; details are kept portable.
AI Assist LLM · RAG
Founded an LLM + RAG system (built on Claude via Amazon Bedrock) that reads a complex FortiCNAPP alert and produces a prioritized incident report and an alert-grounded remediation plan. Set the roadmap at launch and shipped every milestone — through to an alert-context API that plugs into customers' own agentic / MCP workflows. Drove a 30% lift in analyst engagement. Read the blog →
MCP server MCP · Agentic
Designed and shipped the MCP server for FortiCNAPP: the agent-ready interface to live cloud-security data. Customers get prioritized, trustworthy answers inside Claude across alert review, tuning, compliance, and vulnerabilities, with no API key, per-user access, read-only admin-revocable connections, and results that declare when data is partial.
Composite Alerts engine Graph · ML
Founding engineer and primary architect of the correlation engine: entity resolution collapses signals into canonical entities, then an ephemeral Neo4j graph correlates 400+ detections across 140+ MITRE ATT&CK techniques — cloud, host, container, ML-anomaly, small-language-model, and runtime vulnerable-activity (RiskWatch) signals — into single high-fidelity incidents, cutting false positives 40%.
Lakehouse migration Spark · Iceberg
Designed and led the migration of the detection/incident pipeline off a proprietary warehouse onto an open Spark + Apache Iceberg lakehouse, re-architecting a monolith into independently scalable producer/consumer services — saving $1.5M+ in annualized cloud cost. The latest phase runs 40% more detections at 47% lower cost, so adding a detection no longer adds query load.
LLMOps & AI security Eval · Red-team
Built the practice that keeps GenAI trustworthy: an LLM-as-judge evaluation harness, adversarial red-teaming (5,000+ prompts), production SLOs, and per-request cost/latency observability — plus AI-threat detection policies and a reusable AI platform service other teams build on.
Detection models SLM · Classical ML
Co-designed the small-language-model detectors that feed the correlation engine — a command-line anomaly model and a hostname-injection model, both built on fine-tuned transformer encoders (transfer learning) and each pairing a supervised classifier with unsupervised anomaly detection to catch known and novel attacks. I also own a wider fleet of 10+ statistical and ML detectors (GMM, Isolation Forest, ARIMA, NER + gradient boosting) across cloud, host, and identity telemetry, plus the human-in-the-loop alert-scoring loop behind Composite Alerts — cheap, batchable signal that adds coverage the rule-based detectors miss.
Data platform & scale Streaming · Rust · Ingestion
The detection fleet and correlation engine run on top of a large ingestion backbone: a host agent streaming 25+ types of behavioral telemetry, carried over a Kafka-protocol streaming pipeline and decoded by a Rust-based ingestion service that autoscales to hundreds of pods per region — feeding both an hourly batch-analysis tier and a separate, sub-15-minute near-real-time tier for faster detection paths.
Before that — a decade of production ML & data science HPE · Salesforce
At HPE InfoSight (HPE's AIOps platform): workload anomaly detection, a recommendation engine that auto-resolved 86% of customer cases, cross-stack analytics across host / storage / network, and capacity forecasting & sizing. At Salesforce / Tableau: subscription-churn forecasting (−1.5% churn) and NLP for product ideas, plus data pipelines for the Tableau Data Equity Hub.
Design → Outcome
How I work — design first, then ship it.
Two efforts where I wrote the design up front and drove it to production. The point isn't just the result — it's that the architecture I proposed is the one that shipped.
Re-platforming a security correlation engine
From a warehouse-coupled monolith to an open lakehouse — by design
A correlation engine I help found and architect runs 400+ detections an hour — resolving entities and correlating them in an ephemeral Neo4j graph into high-fidelity incidents. It was coupled to a proprietary data warehouse — costly, and hard to scale the ingest and analysis stages independently. I wrote the migration design before we touched code.
◆ The design I authored
- Split the monolith into a producer (pre-aggregates detection "edges" into an open Iceberg lakehouse) and a consumer (graph analysis → alerts) — so each scales independently.
- Single-scan grouped caching: read each source once, derive every policy off the cached base, union, write once — collapsing 155 repeated scans of one table into one.
- Per-pipeline triggering, managed reference-data tables, and metrics parity carried over from the legacy system.
▲ What shipped
- The producer/consumer split landed in production, validated end-to-end on real tenants.
- The single-scan strategy shipped as a one-pass query rewrite across 92 detection policies, byte-for-byte identical output.
- Reference data moved to managed lakehouse tables; observability and alerting reached parity.
Founding an LLM product line
A roadmap set at launch — and every milestone shipped
Analysts were drowning in dense alerts. I founded AI Assist, an LLM + RAG system built on Claude, and laid out a year-long roadmap on day one — then delivered all of it. I wrote about it here →
◆ The vision I set
- Natural-language querying of an alert → structured, prioritized summaries → grounded remediation → an alert-context API for agentic / MCP workflows.
- Built on a reusable platform service, gated by evaluation and red-teaming so it stays trustworthy.
▲ What shipped
- Every milestone, in sequence, over 12 months — through dedicated AI Incident Report & Remediation tabs and the agentic API.
- An LLM-as-judge eval harness, adversarial red-teaming, and production SLOs; adopted by a second product team with zero new infrastructure.
Experience
Where I've worked
Lead a team across detection, GenAI, and platform. Founded AI Assist; founding engineer and architect of the Composite Alerts engine; designed and led the lakehouse migration; own the cross-team delivery roadmap.
Subscription-renewal churn forecasting (−1.5% churn), NLP topic modeling for product ideas, SHAP-based viz-performance models, and census-data pipelines for the Tableau Data Equity Hub.
Built ML for HPE InfoSight (HPE's AIOps platform) — workload anomaly detection, a recommendation engine that auto-resolved 86% of customer cases, cross-stack analytics for virtual environments, and Resource Planner capacity forecasting/sizing.
Predictive models for storage issues — early work in the Nimble lineage that became HPE InfoSight — plus executive dashboards and Flask APIs, improving support satisfaction 12.5%.
Toolkit
How I build
Recognition
Patents, writing & education
- Patent (Issued, USPTO): US 12,627,687 — Generative AI-Based Multi-stage Composite Event Processing — co-inventor.
- Patent (Granted): US 9,065,952 B2 — determining the significance of performance issues in IT environments. Patent (Pending): ML-based detection of compromised cloud credentials. Patent (Filing Approved): LLM-based triage of cloud alerts using historical context and customer-provided guidance.
- Writing: "FortiCNAPP AI Assist: Accelerating the Investigation and Remediation of Cloud Threats" — authored, Fortinet Community.
- Education: MS in Analytics, Arizona State University (GPA 4.0, Beta Gamma Sigma); BTech in Electronics & Communication Engineering, JNTU.