Software Architect · Founder, imake software
I build AI systems that hold up in production, not in demos: agents with typed, permission-gated tool access; MCP servers that let AI operate real systems through the same APIs humans use; and the observability to know when they're right. Below is the public evidence you can check yourself, plus the production work I can't link to because it's internal to an employer.
An MCP server for Claude Code that gives coding agents cross-repo architectural context — answers cross-repo impact questions and flags violations of organizational standards.
View on GitHub →Knowledge management system with MCP integration for Claude: SQLite full-text search, entity extraction, git auto-sync, and a background file-watching daemon.
View on GitHub →CLI bridging Linear and git for Claude Code — ticket management, branch creation, draft PRs, and project context handed directly to a coding agent.
View on GitHub →AI-powered developer-productivity platform integrating Jira, GitLab, and semantic search over vector embeddings. Multi-model Claude architecture with context compaction for production-viable cost. Shipped in 3 weeks — won the company hackathon.
Read the writeup →Built an MCP server enabling Claude Code integration with internal systems, bringing AI-assisted workflows to the engineering organization.
Read the writeup →Internal project command center exposing a 20+ tool MCP server, so AI agents operate real projects and tasks through the same typed API and per-user auth as human users.
Read the writeup →AI chat assistant integrated with the company knowledge base and internal applications. Tool access is gated by per-user permissions, so the agent and the employee it's acting for only reach data they're authorized for.
Retrieval-augmented assistant over PostgreSQL/pgvector inside a CPA operations suite, plus document-understanding pipelines (OCR + LLM extraction) for financial statement import.
Custom observability and monitoring tool built from scratch for a nine-app production portfolio — the production signal behind everything above.
Read the writeup →This work has centered on production tool-use and agent architecture, typed permission-gated access, MCP interoperability, retrieval, rather than formal offline evaluation harnesses. The closest equivalent is production monitoring, LMS Tech Observatory above, and real-usage feedback loops rather than a dedicated eval pipeline. Happy to talk through how I'd close that gap.