I build agentic LLM systems, RAG pipelines, and full-stack platforms — from production websites serving global users to LLM-powered career tools. Currently Analyst, AI & Data at Indiana University. MS CS, IU. Formerly Mercedes-Benz R&D.
An adversarial sandbox where an attacker agent evolves live web exploits and a defender agent authors real AWS WAFv2 rules to stop them — modeled deep enough that the output is one step from a real AWS deployment.
Most "AI + security" demos are a thin wrapper around a toy check. This one models the actual domain: real WAFv2 statement types, real Web ACL Capacity Unit accounting, real sliding-window rate-limit semantics, and AWS's actual sampled-request log shape — so the artifacts the agent produces are structurally the real thing.
12 WAFv2 statement types modeled with AWS's exact JSON one-of shape, evaluated in priority order with correct short-circuit and cross-rule label propagation.
Every rule stages in COUNT mode, replays against legit traffic with deliberate false-positive bait, and an independent check auto-rolls-back any rule that regresses.
A deliberately tight WCU budget forces the defender to consolidate and retire rules rather than stack regex — mirroring real operational limits.
A hand-rolled tool-calling loop — no LangChain or CrewAI — swaps between local Ollama, a deterministic CI backend, or a hosted LLM with zero agent-logic changes.
Agentic AI, RAG pipelines, and full-stack platforms — each one deployed and tested, not left as a notebook.
Attacker agent evolves live web exploits; defender agent authors real AWS WAFv2 rules to stop them. 12 statement types modeled to production fidelity, WCU-budget enforced, guardrails in code — 67% → 0% attack success across 5 rounds, 0% false positives held.
Full-stack healthcare platform with role-based access for patients, doctors, and insurers. Containerized microservices on EC2, S3 for HIPAA-grade documents, 30+ REST APIs secured with 2FA, and event-driven Lambda processing.
Grounds Claude in real job data through a LangChain + ChromaDB RAG pipeline for resume–job matching, with an NLP-based ATS scoring engine driven by vector embeddings and semantic overlap — outperforms keyword-only baselines across 100+ tests.
Multi-agent logistics app: constraint-based route optimization across 10+ stops via Gemini and Google Maps, 100% pickup-before-dropoff adherence. Fully hands-free voice-to-route through the Web Speech API — cuts manual planning time in half.
"What did I even do yesterday?" — solved. A free MCP server that connects Claude to your GitHub activity: ask for your standup and it pulls the last 24h of commits, PRs, reviews, and issues, formatted and ready to paste into Slack. Ships get_standup, get_standup_multi, and get_repo_pulse tools. Direct HTTPS calls to the GitHub API — no proxy, no telemetry. MIT licensed.
I'm looking for Applied AI, AI Software, and Forward Deployed Engineer roles at teams shipping real AI products. If that's you — let's talk.