I work at the intersection of AI/ML, cybersecurity, operational technology (OT), autonomous vehicle systems, and wireless infrastructure (5G/6G/V2X). I am a Senior Software Design Release Engineer at General Motors Advanced Infotainment, Compute & Connectivity in Milford, MI.

I am continuously learning — building agentic AI systems, RAG pipelines, and multi-agent workflows across several programs simultaneously. This repository is a reflection of that ongoing journey, not a finished product.

9-agent multi-agent pipeline for engineering talent acquisition insights. Specialized agents analyze sourcing quality, rejection patterns, interviewer load, offer trends, pipeline health, cost optimization, and real-time market compensation. RAG-grounded via ChromaDB, evaluated by LLM-as-judge, with two published n8n workflows and a Streamlit dashboard. All 20 security risks driven to LOW residual. Runs end-to-end for under one cent per full pipeline run.

Stack: LangChain OpenAI ChromaDB n8n Cloud FastAPI Streamlit Pydantic
Docs: Architecture · PRFAQ · Security Matrix

Benchmarks a customer support ticket classifier across four approaches: zero-shot GPT-4o-mini (92.9%), few-shot GPT-4o-mini (92.9%, 37% faster), few-shot GPT-4o (100%), and LoRA fine-tuning of SmolLM2-1.7B on a T4 GPU (78.6%). Generates real cost, accuracy, and latency data for build-vs-buy decisions. LoRA adapter: 12.6MB vs 3.4GB full model — 270x compression.

Stack: OpenAI API HuggingFace Transformers PEFT/LoRA 4-bit Quantization Google Colab T4

Implements the full LLMOps feedback loop on a signup email agent: inspect baseline behavior, improve prompts, expand golden dataset, evaluate with DeepEval (ExactMatch, GEval, custom LLM-as-judge). LangSmith tracing integrated for online evaluation. 20-scenario golden dataset at 75% pass rate.

Stack: LangChain LangSmith DeepEval OpenAI GPT-4o-mini Python

Personal AI digital twin built with GPT-4o, ChromaDB + BM25 hybrid retrieval, and Gradio. Security-by-design architecture with dedicated input validation, output filtering, and rate limiting. 23 biography sections, 188 retrievable chunks. Deployed to Hugging Face Spaces.

Stack: Python GPT-4o ChromaDB BM25 Gradio Hugging Face
Live: huggingface.co/spaces/LSmithPMP/digital-twin

Node.js MCP server exposing five cybersecurity tools: search_iocs, lookup_cve, threat_actor_profile, map_ttps, get_risk_score. Zod input validation, token-bucket rate limiter, deny-by-default tool set, zero outbound HTTP. Full PRD, Risk Matrix, and architecture diagram included.

Stack: Node.js MCP Protocol Zod TypeScript Rate Limiting

16-node n8n Agent-to-Agent workflow with parallel Weather and Travel agent microservices, weighted scoring, and ranked city recommendations. Full security hardening to all-LOW residual risk. PRD v2.0 with security architecture section.

Stack: n8n A2A Protocol Multi-Agent Weather APIs Python

AI-powered customer retention system with three parallel implementations — n8n workflow orchestration, React demo interface, and Python backend — demonstrating production-grade agent design patterns.

Stack: n8n React Python LLM Orchestration GPT-4o

Voice-enabled RAG assistant with multi-agent orchestration via n8n. Integrates Whisper for speech recognition, Pinecone for vector retrieval, and ElevenLabs for voice synthesis.

Stack: Python Whisper Pinecone GPT-4o ElevenLabs Streamlit

Agentic AI system orchestrating across five structured healthcare datasets via n8n, demonstrating multi-source agent coordination for clinical decision-support workflows.

Stack: n8n Multi-Agent Google Sheets GPT-4o Python

Four-panel ML performance dashboard: Digital Twin Drift & Anomaly Detection, 5G/V2X Network Inference Latency, AV Sensor Fusion & Model Confidence, RAG & Agentic Pipeline Performance.

Stack: Tableau Public ML Metrics 5G/V2X AV Sensor Fusion RAG

Simulation scaffold for dissertation research: evasion-based and perturbation-based attack modeling against ML models embedded in autonomous vehicle OT systems, using CARLA, ROS, SUMO, and OMNeT++ environments.

Stack: Python CARLA ROS SUMO OMNeT++ PyTorch