Ramp Enhances Receipt Processing Through Fine-Tuned LLM Automation

Ramp has leveraged Modal to fine-tune large language models (LLMs) for improved receipt processing, achieving a 34% reduction in receipts necessitating manual handling.

Overview of Ramp

Ramp is redefining the financial operations sector with a suite tailored for CFOs. Their platform combines corporate cards with a variety of financial management functions, such as expense tracking, vendor management, procurement, bill payments, and accounting integrations. This integration aims to streamline the numerous tasks often faced by finance teams, making processes like receipt uploads and vendor payments more intuitive. It’s all about simplifying complex workflows, thereby enhancing both time and cost efficiency for professionals in finance. Given the intricacies involved in managing finances, tools like Ramp represent a significant shift in how organizations can approach their operations.

The Challenge: Customization Limitations with Initial Providers

A key component of Ramp’s suite is its intelligent receipt submission flow, which uses LLMs to convert Optical Character Recognition (OCR) data into structured JSON. Initially, Ramp explored options from providers like OpenAI but found them lacking in customization. This highlights a critical issue in the functionality of many existing AI tools: while they may work well out of the box, they often fail to adapt to specific business requirements or user needs. Concerns about costs, reliability, and data security further complicated their decision-making process. As they considered fine-tuning APIs for open-source models, it became evident that these black-box solutions also had serious limitations in customizability.

Rahul Sengottuvelu, Head of Applied AI at Ramp, emphasized the importance of flexibility: “Finetuning on Modal allows us to implement custom logic and preprocessing. Additionally, being able to train hundreds of models in parallel helps us accelerate our training iteration and tuning.” This realization underscored the need for Ramp to find a solution that provided comprehensive control over their fine-tuning workflows. It’s not merely a matter of executing tasks; having the ability to refine processes iteratively can significantly influence long-term outcomes. This quest for flexibility speaks volumes about the industry's ongoing struggle with adapting AI technologies for specific business contexts.

Implementing the Solution: Enhanced Accuracy and Cost Efficiency

The adoption of Modal enabled Ramp to achieve significant improvements in their receipt processing capabilities. The new infrastructure reportedly cost 79% less than notable competitors like OpenAI, which provided a compelling financial incentive to make the switch. This cost gap is remarkable and raises questions about existing pricing strategies within AI service providers. In an age where financial efficiency is paramount, such disparities could push companies toward newer, more adaptable technologies. As a result, Ramp’s approach led to a swift 34% reduction in receipts requiring manual oversight. This kind of metric isn't just a number; it represents a significant shift in operational workflow.

Modal serves as an adaptable platform capable of running Python functions in the cloud, granting Ramp the necessary flexibility to develop a tailored experimentation framework. The team utilized Modal to:

• Train multiple prospective models concurrently.
• Store weights from different fine-tuning attempts in Modal volumes.
• Establish an inference endpoint capable of deploying various models based on specific inputs.

These functionalities facilitated a swift evaluation of model performance across distinct design options. This capability is essential, as developing a successful AI model often requires a continuous cycle of testing and refinement, one that can be both resource-intensive and time-consuming.

Modal's extensive support has provided additional advantages for Ramp:

• Effortless infrastructure orchestration: Modal's system automatically adjusts GPU resources, addressing a common pain point associated with major cloud providers. Many companies struggle to optimize these resources, often leading to wastage.
• Standardization of experimental environments: The platform allows seamless definition of containerized environments, which can easily integrate into any Modal function. Standardization reduces variability and can improve the reliability of experimental outcomes.

Additional Benefits: Optimizing Batch Processing Efficiency

More than just fine-tuning, the Ramp team identified further applications for Modal's capabilities. The experience raised the question: how quickly can you deploy an application? Chris Nguyen, a Software Engineer at Ramp, remarked, “Just use Modal. You can get an application up in 5 minutes.” This sense of immediacy is invaluable in today’s fast-paced business environment.

A notable challenge involved redacting personally identifiable information (PII) from 25,000 invoices. A task that would traditionally take three days to complete locally was expedited using Modal's functions. By parallelizing this task across 256 cloud workers, the team completed it in merely 20 minutes for about $100. This significant acceleration not only reduces the workload but also mitigates the risks associated with handling sensitive data.

Many organizations are recognizing the importance of efficient infrastructure for data-centric tasks, leading to dramatic increases in developer productivity. Ramp's integration of Modal allows it to enhance AI offerings at a faster pace than ever. The serverless architecture streamlines data processing, which can be both a headache and a bottleneck for many companies, especially as they scale.

Gennady Korotkevich, another Software Engineer at Ramp, reflected on the streamlined experience: “It’s a good feedback loop. It’s pretty easy with Modal.” This highlights an often-overlooked factor in AI solutions—user experience matters significantly. Teams need to adapt quickly, and solutions that are difficult to implement can derail even the best strategies.

Implications and Future Outlook

Ramp’s strategic integration of fine-tuned LLMs through Modal signifies a notable shift in how AI is applied in financial operations. The ramifications extend beyond mere efficiency. Companies that engage with AI for critical processes like receipt handling and data protection can redefine their operational trajectories while concurrently managing risk. What this means for you and your organization is that excellence in operational execution isn't just a goal; it's becoming increasingly attainable with the right tools.

This situation raises a broader question: as other firms look to follow in Ramp's footsteps, will they find equal success? The success here highlights the importance of selecting the right technology and effectively customizing it. The competitive edge is there for those who can implement AI that specifically addresses their unique challenges. Moreover, with growing awareness around data privacy and operational efficiency, the increasing demand for intelligent solutions is clear.

Bottom line: as companies continue to innovate, the prospect of introducing AI-driven tools into finance and other operational domains will likely become more mainstream. This is more significant than it looks. We're witnessing the evolution of finance management—one that may soon dictate standards across various sectors.