[Opening: Introduction & Purpose]
Hello and welcome!
Today, I’ll be presenting the Invoice Reimbursement System, an intelligent
automation platform designed to streamline and standardize employee invoice
reimbursements using advanced AI.
Why did we build this?
Manual invoice review is slow, inconsistent, and often frustrating for both employees
and finance teams. Our system automates this process, ensuring fast, fair, and
transparent reimbursement decisions, and allows anyone to query reimbursement
data using natural language.
[Section 1: High-Level Overview]
This project has two main components:
1. Automated Invoice Analysis:
● Analyzes employee invoice PDFs against the company’s HR
reimbursement policy using a Large Language Model (LLM).
● Stores the analysis results, including invoice details, status, and
reasons, in a vector database for efficient search.
2. Intelligent Query Chatbot:
● Lets users ask questions about processed invoices in plain English.
● Uses Retrieval Augmented Generation (RAG) to find and summarize
relevant invoice data.
[Section 2: Core Technologies Used]
● Python as the main programming language.
● FastAPI for building robust, async API endpoints.
● Large Language Models (LLMs): OpenAI GPT, Gemini, or open-source LLMs
for understanding and analyzing text.
● Vector Store: ChromaDB (or similar) to store embeddings and metadata for
fast, semantic search.
● Embeddings: OpenAI or Sentence-Transformers to convert text into
searchable vectors.
● PDF Parsing: PyMuPDF for extracting text from PDFs.
[Section 3: Workflow – How It All Fits Together]
Let’s walk through the entire workflow:
1. Invoice Analysis (Part One):
● The user uploads:
● The HR policy in PDF format.
● A ZIP file containing one or more invoice PDFs.
● The employee’s name.
● The backend parses the policy and each invoice.
● For each invoice, the system prompts the LLM to decide:
● Is it Fully Reimbursed, Partially Reimbursed, or Declined?
● What’s the detailed reason, referencing the policy?
● The system generates vector embeddings for both the invoice content
and the LLM’s analysis.
● All this data—status, reason, employee, date, and embeddings—is
stored in the vector database.
● The API returns a JSON summary of the results.
2. Chatbot Query (Part Two):
● The user asks a natural language question, like “Show all declined
invoices for John Doe.”
● The system embeds the query and searches the vector store for
relevant analyses.
● It applies any filters (like employee name or status).
● The LLM uses the retrieved documents as context to generate a clear,
markdown-formatted answer.
● The chatbot’s response is returned to the user.
[Section 4: Live Demo – API Endpoints in Action]
Let’s see the system in action using the Swagger UI:
A. Invoice Analysis Endpoint
1. I’ll open the Swagger UI at /docs.
2. I’ll expand the POST /api/invoice/analyze endpoint.
3. Click Try it out.
4. Upload the [Link] and [Link] files.
5. Enter the employee name, for example, “Sunny Raj.”
6. Click Execute.
7. The system processes each invoice, analyzes it against the policy, and returns
a JSON response.
● Here, you can see the status for each invoice—fully reimbursed,
partially reimbursed, or declined—along with detailed reasons.
B. Chatbot Query Endpoint
1. Now, I’ll expand the GET /api/chatbot/query endpoint.
2. Click Try it out.
3. Enter a query like:
“Show all declined invoices for Sunny Raj.”
4. Click Execute.
5. The chatbot uses vector search to find relevant invoices, and the LLM
summarizes the results in markdown format.
● You’ll see a human-readable answer listing all declined invoices for that
employee, with reasons.
[Section 5: Technical Highlights & Prompt Engineering]
● Prompt Engineering:
● For invoice analysis, the prompt instructs the LLM to read both the
policy and invoice, assign a status, and justify the decision with
references to policy clauses.
● For the chatbot, the prompt guides the LLM to use retrieved analyses
as context and generate clear, markdown-formatted answers.
● Vector Store:
● Enables semantic search and fast filtering by employee, date, or status.
● Robust Error Handling:
● The system validates all inputs and provides clear error messages for
missing files or invalid data.
[Section 6: Benefits and Use Cases]
● For Employees:
● Instant, transparent feedback on reimbursement requests.
● For Finance Teams:
● Automated, consistent validation and a searchable audit trail.
● For Auditors and Managers:
● Easy, natural language access to reimbursement data for compliance
and reporting.
[Closing: Summary]
In summary, the Invoice Reimbursement System combines the power of LLMs and
vector search to automate and modernize the reimbursement process. It’s fast,
transparent, and intelligent—making life easier for everyone involved.
Thank you for watching!