Video Transcript & Sections

Problem Space - Cardin

Hello, I'm Cardin, co-founder of TLG, a profitable and rapidly growing cost reduction consultancy established 5 years ago. I met Toby at an All-In Meetup, who invited me here. Tom joined our team at the event.

TLG specializes in serving businesses with Laundry, Linen, and Uniform programs, with annual spends ranging from tens of thousands to tens of millions of dollars.

We've observed that 95% of our new customers face cost issues that violate their supplier agreements. As a result, TLG conducts regular invoice audits and sends cost correction demand letters to suppliers.

However, these ongoing audits are currently the most time-consuming process at TLG and the main obstacle to our scaling efforts.

Demo / Solution - Toby

I'm Toby, the hackathon technical lead. Tom contributed about half of the code I will describe.

To optimize efficiency, we decided to skip audits when they are not required, which accounts for approximately one-third of the time. To achieve this, we developed an LLM pipeline that converts PDF Supplier Agreements and Supplier Invoices into a list of Supplier Agreement Violations. If this list is empty, it indicates that the audit can be skipped.

The tech stack we used for this project includes Colab, Python, ChatGPT API, LangChain, and Camelot.

Effective use of LLMs and prompts & Technical Merit - Toby

The majority of the data pipeline operates on LLM prompts, covering filters, joins, LangChain agent generation, and dynamic code generation for application-specific data transformations. Only a small amount of boilerplate and PDF parsing stages are not implemented as LLM prompts, and some PDF parsing remains to be automated.

Compared to traditional AI techniques, our solution offers advantages such as higher resilience to concept drift, a common challenge in this type of application, and lower implementation and ongoing support costs.

While there may be other unforeseen advantages or drawbacks, we haven't explored them extensively yet.

Impact - Cardin

Although the implementation is proof-of-concept quality, it is a commercially applicable e2e data pipeline.

If scaled to production quality, it would have the following impacts on our auditing process: 30% OPEX Reduction Reduce cycle time from 2 months to instantaneous Reduce customer Refund Days Outstanding from 12 to 3 weeks Increase TLG's customer capacity by X%, TBD

I believe this will also improve customer satisfaction.

Creativity - Cardin

In conclusion, I'm inspired by our achievement because we developed a tool for an industrial niche that lacks access to modern tooling. Currently, there is no existing commercial solution for this application.