AI Build & Implementation
Company X AI Agent
A multi-tool agentic AI system built for an independent restaurant delivery platform — shipping customer support, onboarding, menu writing, and personalization in one interface.
~65%
// support_tickets_automated
6wk
// concept_to_deploy
5
// tools_in_agent
$0
// headcount_added
// the_problem
A platform scaling faster than its operations
Company X was in active pilot stage with a handful of independent restaurants — but had no support, marketing, or onboarding team to serve them.
Restaurant owners lacked the time or skills to write compelling menus, respond to reviews, or navigate their own setup progress.
"Where is my order?" queries were handled manually — consuming 3+ hours per day the lean team couldn't spare.
Scaling to more restaurant partners without solving these bottlenecks would fracture the platform under its own momentum.
// measured_impact
Before & after
Before
3h/day
Manual customer support queries handled by the team
→
After
~12m
Edge cases only — agent handles everything else
Before
3–4 days
To onboard a new restaurant partner to the platform
→
After
<4h
Self-serve onboarding with AI guidance at every step
Before
0%
Owners writing their own menu descriptions
→
After
100%
AI-generated copy, approved in under 60 seconds
// system_architecture
How the system is built
The agent runs on a 4-layer architecture — from user input through Claude's reasoning engine, down to tool execution and secure infrastructure. Each layer is independently testable and replaceable as Company X scales.
Company X Agent — System Architecture
L1
Interface Layer — React + Vite SPA
Chat UI (React)
Message history state
Tool call visualization
Typing indicators
Responsive layout
↓ POST /api/chat → { model, messages, tools }
L2
Reasoning Engine — Claude Sonnet + Tool Use API
claude-sonnet-4
Tool use API
Multi-turn context
Agentic loop
System prompt
stop_reason: tool_use
↓ tool_use blocks → tool execution
L3
Tool Layer — 5 Serverless Functions
check_order_status()
generate_menu_description()
draft_review_response()
get_onboarding_checklist()
get_recommendations()
↓ tool_result → back to L2 context
L4
Infrastructure — Vercel Edge + Secure Proxy
Vercel serverless
API key in env vars
Zero client exposure
GitHub CI/CD
HTTPS everywhere
// tool_definitions
Five tools, one agent
Each tool is defined with a JSON schema that Claude uses to reason about when and how to call it. The agent selects the right tool — or sequences multiple tools — based on the user's request.
01
Support
Order Tracker
check_order_status({ order_id: string })
Returns status, ETA, driver, and items for any order ID. Eliminates the most common support request without human involvement.
02
Marketing
Menu Writer
generate_menu_description({ dish_name, ingredients, cuisine_type? })
Generates compelling dish copy from minimal input. Removes the copywriting barrier for time-strapped owners — especially non-native English speakers.
03
Marketing
Review Responder
draft_review_response({ review_text, rating: number, restaurant_name })
Classifies review sentiment (positive / neutral / negative) and drafts a tone-matched response. Protects brand reputation without owner effort.
04
Ops
Onboarding Status
get_onboarding_checklist({ restaurant_name: string })
Returns completion %, completed and pending steps, and the highest-impact next action. Cuts go-live time from days to hours with zero manual check-ins.
05
Growth
Recommender
get_recommendations({ customer_id, time_of_day, preferences? })
Returns personalized picks based on order history, time of day, and neighborhood trends. Drives repeat orders — the core metric in food delivery economics.
// live_interaction
Multi-turn reasoning in action
The agent handles complex multi-step requests in a single thread — calling multiple tools, feeding results back into Claude's context, and synthesizing a single coherent response.
What's the status of order FH-1042?
▸ tool_use — check_order_status
{ "order_id": "FH-1042" }
✓ tool_result
status: "Out for Delivery" · driver: "Miguel R." · eta: "~8 minutes"
Your order from Casa Verde Kitchen is on its way — driver Miguel R. is about 8 minutes out. Your Carne Asada Plate and Horchata are coming right up.
Draft a response to this 2-star review: "Cold food, arrived 45 minutes late."
▸ tool_use — draft_review_response
{ "rating": 2, "restaurant_name": "Casa Verde Kitchen", "review_text": "..." }
✓ tool_result
tone: "negative" · draft_response: "We're truly sorry..."
"We're truly sorry to hear about your experience at Casa Verde Kitchen. This is not the standard we hold ourselves to — please reach out to support@companyx.com so we can personally make this right."
// agentic_loop
How the loop works
1
User sends a natural language request
POST /api/chat → { messages, tools, system }
No forms or menus. The agent understands intent from plain language and begins reasoning about which tool — or sequence of tools — is required.
2
Claude selects and calls a tool
stop_reason: "tool_use" → content[].type === "tool_use"
When Claude decides to use a tool, the API returns with stop_reason "tool_use". The frontend reads the tool name and input parameters from the response content blocks.
3
Tool executes server-side
executeTool(name, input) → structured JSON result
The tool function runs — querying data, generating content, or executing logic — and returns a structured JSON result that gets appended to the message history.
4
Result feeds back to Claude
{ role: "user", content: [{ type: "tool_result", ... }] }
The tool result is added to the conversation as a tool_result content block. Claude receives the full context and continues reasoning — calling more tools if needed.
5
Claude synthesizes a final response
stop_reason: "end_turn" → content[].type === "text"
When Claude has all the information it needs, it returns stop_reason "end_turn" with a natural language text response that synthesizes everything from the tool calls.
// tech_stack
Technology used
Frontend
React + Vite
AI Model
Claude Sonnet
API Layer
Tool Use API
Backend
Node.js + Python
Data
PostgreSQL + ETL
Hosting
Vercel Edge
Security
Env Var Proxy
CI/CD
GitHub Actions
// outcomes
What this delivered
~65% support automation. The most common customer queries are handled autonomously — freeing the Company X team for growth, not inbox management.
Onboarding from days to hours. New restaurant partners self-serve their entire setup with AI guidance — no manual check-ins required.
Enterprise-grade tools for indie owners. Independent restaurants now have AI-powered menu copy and review management on par with large chains.
Production-ready from day one. Deployed on Vercel with a secure API proxy, polished UI, and zero exposed credentials. CI/CD on every push to main.
Architected to extend. New tools can be added by defining a JSON schema and a serverless function — no changes to the core agentic loop required.
Want something built like this?
Khoda Consulting designs and ships AI agents, data pipelines, analytics tools, and ML solutions for growing businesses.