I build AI agents.
Then I build what stops them going rogue.
I design production multi-agent systems with LangGraph and AWS Bedrock — then bring an AWS Security Specialty and CEH background to harden them against prompt injection, tool misuse, and tenant data leakage before they ship.
Security depth meets agentic engineering
I'm a cybersecurity graduate student at the University of Maryland who spends most of my time somewhere most security people don't: inside the agent loop itself. I build the multi-agent systems — orchestration, retrieval, tool use — and then I build the guardrails that keep them from misbehaving in production.
That combination started at Kodryx AI, where I mapped LLM attack surfaces and shipped detection rules against real alert volume, and it sharpened at Tesslate, where I shipped full-stack features for a production AI platform and built the runtime tool-misuse guardrails that keep its coding agent in check, aligned to the OWASP Agentic Security (ASI02) risk class.
I'm not interested in security bolted on after the fact. I want to be in the room while the agent architecture is still being decided — because that's the only place these controls actually hold.
What I actually build with
Split the way the work splits: building the agent, then defending it.
Agentic Systems
GenAI & Retrieval
AI Security & Guardrails
Detection & Monitoring
Cloud & Infra
Languages & Frameworks
Where I've shipped this
● Remote
Agentic Software Engineer Intern
- Shipped full-stack features end to end across a Next.js and React/TypeScript operator interface and a FastAPI orchestration backend, deployed via Docker and Kubernetes.
- Integrated the orchestration layer with internal and external systems over MCP servers and REST APIs, unifying fragmented data sources behind one consistent service interface.
- Root-caused and resolved 6 production defects spanning ORM relationship-loading, frontend error handling, and orchestration routing, cutting support escalations 25%.
- Built runtime guardrails for the AI coding agent — prompt-injection detection, tool-call budget enforcement, loop detection, dangerous-sequence blocking — aligned to OWASP ASI02 and adopted across 8+ agent workflows.
Remote
Security Researcher
- Automated SAST/DAST triage with a Python pipeline wired into Snyk and ServiceNow, cutting review time 40% with no added headcount.
- Mapped 15+ LLM attack vectors — prompt injection, adversarial input, data leakage — and shipped guardrails that closed them across the org's AI feature pipeline.
- Authored 5–10 detection rules across Splunk, ELK, and Chronicle SecOps mapped to MITRE ATT&CK, triaging 200+ monthly alerts and lifting detection accuracy 18%.
- Served as the team's security point of contact, translating findings into CVSS-prioritized remediation roadmaps and raising overall security posture 30%.
Three systems, three failure modes closed
The featured builds below all share a shape: an agent that could do real damage, and a control that stops it.
Secure Multi-Tenant Agentic RAG
Sept – Dec 2025- Built an agentic RAG pipeline on LangChain + AWS Bedrock, scoping every agent's retrieval context to the authenticated tenant's namespace to block cross-tenant leakage.
- Propagated JWT-based tenant identity through Cognito and API Gateway, enforcing org-level isolation across FastAPI routing, per-tenant S3 buckets, and Pinecone namespaces.
- Logged every query, retrieval, and agent response end-to-end via DynamoDB and CloudWatch for compliance-grade forensic traceability.
Agentic Healthcare Prior Authorization Copilot
Mar – May 2026- Designed and ran a 7-agent LangGraph system automating prior authorization — FHIR-aware extraction, insurer-criteria gap analysis, and risk-scored auto-appeal generation.
- Built a hybrid retrieval pipeline (BM25 + Pinecone, BGE/Cohere reranking) with grounded citations and tiered memory, then fine-tuned Llama 3 via QLoRA on approvals data.
- Instrumented hallucination rate, evidence completeness, and Recall@K/MRR through LangSmith and OpenTelemetry, with human-in-the-loop checkpoints on every high-risk call.
Tool-Misuse Guardrails for Agentic AI
May – Jul 2026- Designed runtime guardrails for a production AI coding agent — prompt-injection detection, tool-call budget enforcement, loop detection, and dangerous-sequence blocking — aligned to OWASP ASI02.
- Integrated agent tooling with enterprise systems via MCP and REST APIs across a Kubernetes-deployed FastAPI backend.
- Documented guardrail policies and operational runbooks to make safe agent execution repeatable across the team.
LLM-Augmented SOC Triage Engine
Routed normalized Splunk alerts through Claude API for AI-assisted triage, then caught a high-confidence hallucination via engineered failure tests — proof that model confidence isn't analyst accuracy.
FlowPay — Cloud-Native Payment Orchestration
Payment platform handling authorization, capture, settlement, refunds, reconciliation, and webhooks through event-driven sagas with rollback and DLQ-backed SQS — so a duplicated event never leaves order state wrong. Built with Claude Code.
Data Trustworthiness in Mobile Crowdsourcing
GAN + Random Forest pipeline to flag malicious crowdsourced contributors in real time; synthetic augmentation lifted classification accuracy from 78% to 99.4%.
Testing whether LLMs can be trusted with the alert queue
EAST-LLM: Evidence-Aware Architecture for Automated Alert Parsing & LLM Triage
- Architected a parser cascade normalizing heterogeneous alert formats (JSON, syslog, CEF/LEEF, Suricata EVE, Wazuh) into a unified evidence schema before any model inference runs.
- Built a prompt-guided LLM layer that emits strict-schema, MITRE ATT&CK-aligned classifications — separating deterministic parsing from semantic reasoning by design.
- Hardened outputs with an evidence-aware validation engine enforcing schema, taxonomy, and prompt-injection guardrails, evaluated across CICIDS2017, UNSW-NB15, and Splunk BOTS on F1 and unsupported-claim rate.
Verified skills
Master of Engineering, Cybersecurity
B.Tech, Computer Science (Cybersecurity)
- Program Representative, Cybersecurity Engineering — UMD Graduate Student Government
- 1st Place, Green & Sustainable Project Development Contest — IoT-based Smart Blind Stick
- Winner, 48-Hour Idea Sprint Challenge in Machine Learning — GCGC, GITAM University
Building or securing an agent? Let's talk.
Open to full-time roles in agentic AI engineering, AI/LLM security, and cloud-native security engineering.