Homeland Itoya Events Assistant

Stack

• TypeScript
• Express
• Prisma
• PostgreSQL
• LLM tool-calling (Gemini / Claude / OpenAI)
• Structured agent loop
• React (Vite)
• Tauri
• Next.js
• REST APIs
• optional pgvector RAG

The Operational Challenge

Event teams juggle leads, proposals, vendors, schedules, and client follow-ups across chats, spreadsheets, and tribal knowledge. Questions that should take seconds — “what’s in the pipeline?” or “refine this proposal budget” — meant jumping between tools and threads.

Homeland Itoya Events needed one workspace that covered the planning lifecycle, plus an assistant that could answer in natural language and take structured actions (read metrics, draft proposals, refine contracts, send follow-ups) — without a separate UI for every step.

The Architecture Stack

The system centers on a TypeScript/Express API that runs an LLM agent orchestrator with explicit tool-calling: the model proposes typed tools, the server executes them against domain actions (leads, events, proposals, contracts, emails, vendors, tasks, analytics), and results feed back into the conversation in a controlled, auditable loop — including write confirmation for mutations.

Staff work in a Tauri desktop client (React/Vite) against that API. Clients review proposals and chat through a separate Next.js portal. Business logic stays on the shared backend so web portal and desktop don’t diverge.

• Agent service: structured tool catalog, planner/executor loop, knowledge base + optional RAG
• Desktop (Tauri): day-to-day ops UI — dashboard, lifecycle, proposals, contracts, vendors, analytics
• Next.js portal: secure client review, documents, and AI-assisted Q&A
• Domain tools: leads/events, proposal & contract generate/refine, email automations, vendors/inventory/tasks, analytics snapshots

Reliability & Agent Design

Agent systems fail when tools are vague or side effects are uncontrolled. This assistant uses typed, purpose-built tools and a structured loop so the model requests actions instead of inventing free-form database writes. Writes that mutate documents use confirmation tokens and audit trails.

Thin clients (desktop + portal) share one API brain: ops run in a native app; clients use the portal; both hit the same automation core.

• Tool-calling over free-form replies for operational actions
• Shared Node API for desktop + portal to avoid divergence
• Production deployment with ongoing maintenance for the events team

The Concrete Outcome

Delivered to production from November 2025 and actively used by the Homeland events team — end-to-end LLM agent product work: model orchestration, APIs, React/Tauri ops UX, Next.js client portal, and desktop packaging.