Senior Frontend Engineer → AI Full Stack Engineer
16 years experience · Scottsdale, AZ · Open to opportunities
I'm going deep on agentic AI engineering — not just prompting, but building agents from scratch. Tool use, RAG pipelines, MCP servers, human-in-the-loop patterns, and multi-agent orchestration.
| Project | What it does | New concept |
|---|---|---|
| 📅 Calendar Agent | Schedules meetings from natural language | Agentic loop + tool use |
| 🛠️ IT Support Agent | Diagnoses and fixes VPN + tool access | RAG + parallel tools |
| 🔍 Site Search Agent | Semantic product search — finds "running shoes" from "footwear for jogging" | Real embeddings + pgvector |
| 🍽️ Restaurant Agent | Food recommendations with all production patterns | MCP + streaming SSE + Redis cache |
| 🔌 Restaurant MCP Server | Standalone MCP server — any AI connects via SSE URL | MCP protocol + Zod validation |
| 👥 HR Onboarding Agent | Onboards employees — pauses for manager approval, resumes via webhook | Human-in-the-loop + persistent state |
| 🔍 Code Review Agent | Multi-agent code review — Security, Performance and Style agents run in parallel | Multi-agent orchestrator pattern |
| ⚙️ AEM Knowledge Agent | Internal KB agent — answers AEM questions, finds page owners, troubleshoots errors | Hybrid search (SQL + pgvector) + Redis cache |
| ☁️ Document Intelligence Pipeline | Upload PDF → S3 → Lambda → Textract → Bedrock Claude → DynamoDB → Q&A | AWS serverless + Docker + CI/CD |
✅ Agentic Loop while(stop_reason === "tool_use")
✅ Tool Use LLM decides, your code executes
✅ Parallel Tool Calls Multiple tools in one LLM turn
✅ RAG (keyword) JSON search before acting — Project 2
✅ RAG (semantic) Voyage AI embeddings + pgvector — Project 3
✅ Cosine Similarity Vector distance for ranking
✅ MCP Server Tools exposed via standard protocol
✅ MCP Client Agent connects to MCP via SSE URL
✅ Streaming SSE Tokens appear live in React UI
✅ Redis Caching Repeated searches served instantly
✅ Rate Limiting Bottleneck for Anthropic API
✅ Zod Validation Safe tool inputs — no LLM hallucination
✅ Human-in-the-loop Agent pauses → webhook → resumes
✅ Persistent State Full session saved to Neon DB
✅ Idempotent Tools Safe to retry — never duplicates work
✅ State Machine started → waiting_approval → completed
✅ Session Logging Full request/response + token cost
✅ Multi-Agent Orchestrator + 3 parallel specialists
✅ Promise.all() All specialist agents run simultaneously
✅ Hybrid Search SQL keyword + pgvector semantic combined
✅ Configurable Search Runtime switch between search modes
✅ SQL KB Pattern LLM queries structured DB via tools
✅ AWS Lambda Serverless event-driven functions
✅ S3 Event Triggers Upload → auto-trigger Lambda
✅ Docker + ECR Containerized Lambda functions
✅ CI/CD GitHub Actions → ECR → Lambda deploy
AI
Anthropic Claude · Voyage AI · MCP Protocol · RAG · pgvector · Streaming SSE · AWS Bedrock · AWS Textract
Frontend
React 18/19 · TypeScript · Redux Toolkit · TanStack Query · Vite · AWS Amplify
Backend
Node.js · Hono · Express · Cloudflare Workers · AWS Lambda · Zod · Bottleneck
Data
Neon Postgres · Drizzle ORM · Upstash Redis · pgvector · AWS DynamoDB · AWS S3
DevOps
Docker · AWS ECR · GitHub Actions · AWS API Gateway · AWS CDK
2010 → PHP, Perl, jQuery — e-commerce foundations
2015 → React, Redux — modern frontend
2020 → TypeScript, micro-frontends, design systems
2026 → AI agents, RAG, MCP, AWS serverless, agentic engineering
Key work: Migrated 150+ retail sites (Bazaarvoice → PowerReviews) and LivePerson migration. Oracle Endeca search integration, Adyen payments, Perl automation eliminating manual regional rollouts across large e-commerce program.
maheshbuilds.dev — case studies, blog, and AI engineering notes