Projects
Things I've built
Spring Boot MCP Agent
A full-stack agentic developer tool: a Java 21 / Spring Boot backend publishes tools over an MCP-style producer surface (tools/list, tools/call) and runs a streaming agent that consumes them, with a React + TypeScript front-end that renders tool calls and tokens in real time over SSE. Includes an OpenAI-compatible /v1/chat/completions endpoint and a JUnit suite covering the tools, registry, agent logic, and controllers. The LLM seam is isolated so a real Claude tool-use call drops straight in.
MCP Integration Server
Enterprise MCP server exposing 6 agentic tools over both stdio and Streamable HTTP transports, built as both a producer of tools and a consumer of the protocol. The live demo runs a Claude agent as a real MCP client: it discovers tools at runtime over the protocol, then calls them to handle customer lookups across 5 accounts, surface support tickets, search a 5-document knowledge base, and retrieve live business metrics ($5.8M ARR, 1,340 accounts).
Streaming Analytics Dashboard
A full-stack Next.js dashboard: MRR/ARR, churn, engagement, and segment views, with two real Claude integrations on top: a natural-language query interface that streams answers token-by-token over the live metrics, and server-side cached insight cards. Real-time streaming and a polished React surface over a production data layer.
Portfolio Analyst Sub-Agent System
A Python lead orchestrator (Claude Sonnet 4.6, adaptive thinking) dispatches to three specialist sub-agents (Claude Haiku 4.5) via tool-use (market data, news sentiment, risk concentration), then synthesizes their briefings into a decision-grade analyst memo. Backed by a Claude-as-judge evaluation harness scoring routing, coverage, and quality across hand-written cases. Prompt caching on every system prompt and the lead's tool definitions; cache hit/miss telemetry surfaced in the UI.
HR Operations Agent
An AI-native take on HR operations: an agentic workflow on Claude with real tool-calling, where the interface is driven by the underlying employee data rather than static layouts. The agent orchestrates 8 tools across 30+ employees, 6 departments, and 7 policies, handling compensation analysis, retention risk scoring, org-chart traversal, and PTO tracking in under 3 seconds per query. Full-stack Next.js / TypeScript, live and open-source, a data-driven enterprise interface built end to end.
RAG Knowledge Agent
End-to-end RAG over a documentation corpus: TF-IDF cosine retrieval (smoothed IDF weighting, cached per-chunk vectors) surfaces the top-k chunks with relevance scores, then Claude composes a grounded answer with inline source citations. The retrieve-then-generate pattern behind enterprise knowledge assistants. Both halves are real, no mocks.
CI/CD Pipeline Optimization
Parallelized and cached GitLab CI/CD pipelines for a federal enterprise client, reducing deployment time by 40%. Tightened the loop between a code change and a deployable build, enabling faster iteration on features and prototypes.
Past work · no public repo