Forward Deployed Engineer (Gen AI)
2025 – Present
PwC · Tampa, FL
Multi-agent threat-intelligence platform — Designed and shipped a production system that autonomously maps an organization's external attack surface (subsidiaries, domains, subdomains, IP ranges/ASNs, open ports), replacing manual reconnaissance with an LLM-augmented agent swarm.
Architected it as containerized microservices — a FastAPI supervisor orchestrating FastStream worker agents over RabbitMQ RPC, with PostgreSQL + pgvector and Redis — and designed a resilient dispatch layer that runs the identical pipeline distributed or in-process as a fallback, so a broker outage degrades gracefully instead of hard-failing the scan.
Built a deterministic-tools-first, LLM-as-gap-filler pattern with strict JSON-schema output, a JSON-repair layer, and Langfuse tracing/evals; streamed live progress to a Vue 3 SPA via SSE over Redis, and shipped to Kubernetes via Argo CD GitOps with a GitHub Actions pipeline building 7 services in parallel across four environments.
AI-assisted risk & compliance service — Built a FastAPI + MongoDB microservice that grounds risk and control generation in NIST 800-53 (~1,196 controls) using a deterministic-retrieval → LLM-judgment → human-in-the-loop pipeline, keeping every AI output auditor-defensible.
Developed a multimodal evidence-review pipeline (vision models for screenshots, text extraction for documents) that rates control effectiveness on a 0–100 scale, backed by an offline eval harness scoring AI output against ground truth; exposed the REST surface as MCP tools so the same functions power both the product and a tool-calling agent.
LLM access-review system — Built a LangChain/LangGraph workflow that ingests user-access signals, classifies security risk with AI-generated reasoning and deterministic fallback rules, and surfaces results to managers — owning the full stack from data pipeline through LLM integration and dashboard.