chayanchauhan/ParcelPilot_Assistant

ParcelPilot Support Chatbot

This is my submission for the CalQuity AI Engineer assessment.

I built a customer-facing support chatbot for ParcelPilot using Streamlit and OpenAI's gpt-4o-mini model. Customers can log in (demo login, picks one of the 4 accounts from the dataset) and ask questions about their orders, cancellations, service credits, and policies. It can also raise an escalation to the human support team, but only after the customer confirms.

I also added a small internal "Ops Radar" page (Problem 1 from the extra credit section - proactive issue detection) which shows tickets close to breaching SLA, similar/duplicate tickets, matches to known product issues, and a category-level check for recent ticket-volume spikes.

Every answer in the chat is also tagged with a short reliability badge (high/medium/low) based on what the tools actually returned for that turn - e.g. whether it was backed by a signed customer agreement, touched a deprecated document, or was missing a fact it needed.

How to run this

1. Install the requirements:

pip install -r requirements.txt

2. Add your OpenAI API key. You have two options:

   • Set it as an environment variable:

     export OPENAI_API_KEY=sk-xxxxxx
     

   • OR copy .streamlit/secrets.toml.example to .streamlit/secrets.toml and put your key there instead.

3. Run the app:

streamlit run app.py

4. It will open in your browser (usually http://localhost:8501)

Model used is gpt-4o-mini. I picked this one because it's cheap and fast and handles tool-calling pretty well, which is what this project needed.

What's in each file

• app.py - the streamlit UI (account picker, chat page, Ops Radar page)
• agent.py - talks to OpenAI, runs the tool-calling loop
• tools.py - the actual tools the AI can use (search docs, get data, calculate, escalate)
• data_store.py - loads the excel data + the fee/credit calculation logic
• documents.py - loads and searches the PDF documents
• insights.py - the "Ops Radar" proactive detection logic (SLA risk, duplicate clusters, known-issue matches, volume spikes)
• data/ - the pdfs and excel file given in the assessment pack

Login

On the login screen, pick any of the 4 customer accounts from the dropdown (Northstar Logistics, LumenWorks, Beacon Retail, Axis Labs) and hit Continue. This is a fake/demo login since building real authentication wasn't the point of this assessment.

More details about how I designed things is in ARCHITECTURE.md and PRODUCT.md.