Case Study · Customer Experience × AI

Teaching the business to hear half a million customers.

An LLM-powered pipeline that read, classified and audited every support ticket at Caterpillar Signs — then turned the noise into a live dashboard the whole company could act on.

Apoorav Rao/Gen & Agentic AI Intern/Jun 2025 – Mar 2026
500K+Tickets classified
Faster auditing
15%Fewer tickets / 6 mo
LiveStill in production
The Company

Caterpillar Signs,
part of Group Bayport.

A global, online-first custom-print manufacturer headquartered in Atlanta — customers design products online, and the company prints, fabricates and ships them worldwide. I worked from the Gurugram, India office.

  • Custom everything — signage, banners, covers for furniture & vehicles, décor, and trade-show displays.
  • Six countries across the US, UK and India, shipping via FedEx & UPS.
1,450+
Employees
6
Countries
3
Continents
Family of Brands
BannerBuzzCovers & AllVivyx PrintingCircle OneGiant MediaNeon Earth
My Role

Embedded inside the customer,
not the codebase.

I joined as a Gen & Agentic AI Intern inside the Customer Experience team — not on the engineering floor. That meant I lived with the support and operations problems firsthand, long before I built anything to fix them.

Jun 2025
Joined CX
Mar 2026
10-month build

"I got to see the problem before I was allowed to build the solution."

Sitting in the customer-facing function meant every technical decision started from a real, observed pain — not a ticket spec handed down a chain.

The Problem

Auditing tickets by hand
couldn't keep up.

The CX team audited resolved tickets to find recurring pain points and feed fixes back to product, logistics and tech. But the entire audit was manual — and that capped how much the business could ever learn.

  • A massive bottleneck against the company's real ticket volume.
  • Insights stayed buried in unstructured text nobody had time to read.
  • Recurring issues slipped through — fixed one ticket at a time, never at the root.
150200

tickets per day — the hard ceiling for a single analyst reading them by hand.

What I Built

An end-to-end pipeline that audits
every ticket — automatically.

01INGEST
Pull resolved tickets
A batch job collects every newly-resolved ticket on a schedule — migrated live from Freshdesk to Salesforce mid-project when the company switched CRM.
Salesforce APIFreshdesk APIevery 30 min
02CLASSIFY
Read & understand
Each ticket runs through GPT-4o with zero-shot classification — no training data. It categorises the issue and extracts the product, the exact question, and the resolution.
GPT-4oLangChainzero-shot
03STRUCTURE
Model & store
Outputs are transformed with dbt and loaded into a cloud database. Apache Airflow orchestrates the whole flow — retries, backfills and dependencies, hands-free.
dbtPostgreSQLAWS RDSAirflow
04SURFACE
A live dashboard
Clean, classified data flows into a real-time dashboard the CX team can query themselves — turning a backlog of text into answers anyone can act on.
Zoho Analyticsreal-time
The Final Product

What customers actually struggle with — at a glance.

Interactive · click any product line
S
CX Signal
Support Ticket Intelligence
LIVE · auto-refresh 30 min
500,247 tickets classified
Product Line
click a row
Custom Covers
Most-asked questions
01
How do I measure my furniture for an exact fit?
Usage
21,400
02
Will this cover hold up in winter and heavy rain?
Product
17,800
03
What's the difference between the fabric weights?
Product
13,200
04
My cover arrived without the tie-down straps.
Tech
9,600
05
Can I reorder the same custom size as last time?
Usage
8,100
Query Mix
38%top type
Product31%
Usage38%
Tech11%
Shipping14%
Billing6%
Pain points surfaced
High impact
Sizing & measurement confusion drives pre-sale questions
34% of pre-sale contacts
Medium impact
Tie-down hardware reported missing on arrival
recurring fulfilment gap
Medium impact
Weather-durability expectations unclear at checkout
product-page signal
The Impact

What changed for the business.

500K+
support tickets classified during the internship.
1,200
tickets processed per hour — versus 150–200 per day, by hand.
jump in auditing efficiency — a 400% improvement, running continuously.
20%+
more recurring issues surfaced that manual auditing had been missing.
15%
drop in overall ticket volume over six months — problems fixed at the root.
In production
the pipeline is still running and auditing tickets today.
From Insight to Action

A finding only counts
if someone acts on it.

Half the job was translation — turning raw classified data into plain insights the tech, production, logistics and CX-leadership teams could actually implement. One real loop:

Signal
Classification shows customers repeatedly confused by one product's installation steps.
Flag
I raise it to product & design with the exact questions and volumes attached.
Fix
Clearer instructions and product-page guidance ship to address the root cause.
Result
Fewer repeat tickets on that product — part of the 15% volume drop.
Under the Hood

The stack, end to end.

Ingestion

Salesforce API
Freshdesk API
REST APIs
FedEx / UPS API
Batch processing

AI Layer

GPT-4o
OpenAI API
LangChain
Zero-shot classification
Structured extraction

Data

dbt
transform
PostgreSQL
AWS RDS

Orchestration

Apache Airflow
Retries & backfill
Dependency mgmt

Delivery

Zoho Analytics
Live dashboards
Python
glue & logic
Thank You

Let's build things that listen to customers.

www.apoorav.online
Apoorav RaoGen & Agentic AI Intern500K+ tickets · faster · 15% fewer tickets