Every message enters through a single LangGraph StateGraph. A supervisor node uses structured output to classify intent and route to exactly one of six specialists, rather than letting one model juggle every tool.

Conversation state persists through a MemorySaver checkpointer, so each turn carries context without a database round-trip. Separating routing from reasoning keeps the control flow inspectable, you can see precisely which agent handled a request and why.

Inventory, meal plan, recipe, nutrition, profile, and shopping each run as an isolated create_react_agent with a scoped system prompt and a small, purpose-built tool set.

Tools are built per request through factory closures that capture the authenticated user, so an agent can only ever touch the right person's pantry. Adding a capability means adding an agent and a tool factory, never editing a monolith.

Ingredient names are messy. People type "tomatoe", "tomatos", or reach for a regional word entirely. Fuzzy matching runs on every lookup, so near-misses collapse onto the same item without exact strings.

Inventory is append-only: items are never hard-deleted, only deducted or marked consumed. That preserves history and sidesteps the race conditions that plague concurrent quantity edits.

The shopping list is deliberately not an LLM. A rule engine applies per-unit thresholds and cross-checks current stock against planned meals to emit out-of-stock, low-stock, and missing-for-the-week items, with zero chance of a hallucinated entry.

Nutrition logs are scored against personal daily targets and roll up into macros, calories, and weekly trends the user can query in plain language.