AI agent case

CARMA AI: a parts selection agent for Dubai

For carma.team we built an AI agent for the first layer of parts requests. It accepts a VIN, photo or car parameters, asks clarifying questions, checks catalog access, finds applicable OEM numbers and prepares a structured answer for the customer or manager.

Project facts

Not a generic chatbot, a checked parts workflow

The agent does not guess the part. It follows a controlled process: user status, catalog access, vehicle identification, part clarification, OEM lookup and handoff to a human specialist when the request becomes complex.

VIN and AUTO flows

VIN/Frame lookup, OCR from photos and parameter-based vehicle selection by make, model, year and engine.

OEM lookup

After the vehicle is identified, the agent moves from part name to applicable OEM numbers and diagrams.

n8n orchestration

n8n controls the scenario branches, webhook triggers and calls between the chat, backend and external APIs.

Django + Redis

The backend handles service logic, tokens, logs and responses, while Redis stores session state and cached catalog data.

Business problem

Primary parts requests take time from managers

A customer rarely sends a perfect part number. They may describe a part in their own words, send a VIN, upload a photo or ask for a cheaper analog. A manager then has to identify the vehicle, check access, find OEM numbers, review compatibility and prepare the next step.

Free-form request

The customer can send text, VIN, photo or incomplete vehicle data instead of a clean catalog query.

Clarifying questions

The agent asks for missing details such as front/rear axle, side, model, year or engine.

Access checks

Before catalog work, the scenario can check user status, balance, tariff and remaining catalog requests.

Manager handoff

If the case is complex, the manager receives the vehicle, request, OEM result and dialog context.

How it works

The agent follows a parts-selection sequence

CARMA AI is designed as a process, not a free-form answering bot. Each step narrows the request and makes the next tool call more reliable.

1. User and access

The scenario checks authentication, verification status and catalog request limits before spending external catalog calls.

2. Vehicle identification

VIN, photo OCR or make/model/year/engine parameters are used to identify the right vehicle modification.

3. Part clarification

The agent interprets plain language, handles slang and asks up to several clarifying questions when needed.

4. OEM result

The backend uses catalog endpoints to find OEM numbers, applicability and links to diagrams or groups.

5. Cache and logs

Repeated data can be stored for 24-30 days, reducing repeated API calls and making support easier.

6. Answer or handoff

The user receives a structured answer, or the manager receives a prepared request with context.

Demo scenario

A typical dialog

The public case does not expose private catalog data, so the scenario below shows the logic rather than a production transcript.

Customer: “VIN: WAUZZZ4M5HD000000. I need front brake discs.”

Agent: identifies the vehicle, checks whether the user has catalog access and asks whether the request is for front axle, rear axle or a full kit if needed.

Backend: passes the request through vehicle lookup, part clarification, OEM search and cached catalog data where available.

Result: the customer receives applicable OEM numbers and the manager receives the vehicle, request, selected modification, found part data and dialog context.

Our scope

What AlphaDog worked on

This case should be read precisely: we are not claiming that AlphaDog built the whole CARMA business platform. The focus of the case is the AI-agent layer and the workflow around primary parts requests.

Conversation design

We designed how the agent starts the dialog, collects missing vehicle data, asks clarifying questions and avoids unsupported answers.

Scenario logic

We mapped VIN-flow, AUTO-flow, image/OCR input, free-text requests, catalog access checks and manager handoff.

API orchestration

The workflow connects the chat layer with backend services and catalog endpoints for vehicle, OEM and user access data.

Reliability boundaries

The agent follows a process with defined states and fallbacks, so complex or uncertain requests can be moved to a human specialist.

Architecture

Workflow logic plus AI clarification

The technical stack is practical: n8n for scenario orchestration, Django for backend services, Redis for state and cache, and external catalog APIs for vehicle and part data.

n8n workflow layer

Webhook triggers, scenario branches, calls to backend services and controlled transitions between user flows.

Django backend

Bot response handling, API endpoints, token work, logging and service code that supports the AI flow.

Redis and Postgres

Session state, cached catalog data, logs and operational records for support and debugging.

Catalog API integration

Vehicle lookup by VIN, parameter-based vehicle lists, OEM search, catalog access checks and user profile data.

Why not a regular bot

The value is in controlled tool use

A regular FAQ bot can answer from a knowledge base. CARMA AI has to interact with account state, vehicle data, catalog availability, OEM search and dialog memory. That requires workflow control and backend logic.

Regular FAQ bot

Good for static answers, company information and simple support questions. It usually cannot reliably operate catalog states or user access rules.

Human specialist

Still needed for complex edge cases, final commercial decisions and unusual requests. The agent prepares context before the specialist joins.

Result

AI handles the first layer before the human specialist

CARMA received an AI layer for primary parts consultation. It helps customers formulate the request, moves through catalog logic and gives the team a more structured handoff instead of a raw message.

VIN or photo input

The agent can start from a VIN string, uploaded photo or car details.

OEM and compatibility context

The result is connected to the identified vehicle, part request and catalog path.

Less manual qualification

Managers join later in the process, when the request already contains useful context.

Chat or voice potential

The same logic can be adapted for website chat, messenger flow or voice intake when needed.

FAQ

Important notes

These details make the case clearer for companies considering their own AI agent.

Did AlphaDog build the entire CARMA platform?
No. The case focuses on the AI agent and workflow layer for primary automotive parts requests inside the existing CARMA context.
Can the same logic work for voice?
Yes. The same scenario can be adapted for voice intake when the business needs phone-style qualification. The exact stack depends on latency, language, CRM and call routing requirements.
What makes this harder than a simple chatbot?
The agent must keep state, check user access, identify vehicle modification, clarify the part, call external catalog APIs and pass structured context to a human when needed.
Can this be adapted to other industries?
Yes, when the company has repeated first-line requests that require data lookup, clarification and handoff: spare parts, equipment, logistics, bookings, service diagnostics or technical support.
Project review

Tell us what you want to build

Send a short message. We will review your task and suggest the right website, CMS, SEO, AI agent or automation path.

We usually reply during the business day.