Case Studies

Below is an overview of select past projects.

Automated Insurance Quotes

Animal insurance marketplace in Europe

Problem

Our client was looking to implement an AI agent to give automated quotes for animal insurance to leads from social media.

Solution

We storyboarded the entire journey of customers interacting with the insurance AI Agent. We then implemented and deployed a robust agentic workflow using RAG and LangGragh. We utilized a tree-of-thoughts prompting technique to eliminate hallucinations.

Outcomes

The client achieved 50% cost reduction and 70% speed compared to previous solutions to quote for animal insurance products.

AI Model Evaluations System

US-based growth-stage real estate tech company

Problem

LLM Evaluations are hard. How can you tell if your chatbot gives out reliable responses? Out of the box solutions did not work for our client as they deal with sensitive personal information.

Solution

We advised the client on best practices with tooling & delivered a customized LLM evaluation suite to benchmark their prompts & LLMs with over 10 different metrics, across saliency, robustness and accuracy, among others. The architecture is event-driven, meaning it can run batch processing both in development and production environments.

Outcomes

Since our client operates in the highly regulated mortgage industry, an internal LLM evaluations protocol increased their confidence in rolling out their customer support AI agent in the highly regulated mortgage industry.

Self-healing website scanning

US-based real estate company

Problem

When our client approached us, they were manually copying and pasting data from over 400 sites of luxury apartments websites into a spreadsheet daily. The sites change very frequently, so hiring a developer to automate this with scraping did not work.

Solution

We leveraged LLMs to write self-healing extraction scripts to automatically extract data from these websites several times per day. We provided an end-to-end extraction pipeline running seamlessly, with observability included.

Outcomes

The client now does not have to spend 40 hours a week updating the data manually.

Model Finetuning on internal data

Multinational e-commerce company

Problem

Localization of e-commerce content is labor-intensive, especially when dealing with a specific product catalog where out of-the-box translation services and tools fall short, especially in "long-tail" languages.

Solution

We fine-tuned LLMs on internal localization datasets to improve translation of B2B product descriptions. This included the tonality and slight cultural differences which were desired in the final copy. We iterated on the solutions with a human in the loop until the translation quality was nearly matching that of human translators. We delivered an internal tool to label the quality of translations to further improve the models.

Outcomes

We cut down the time needed to perform localization of marketing content from hours down to minutes with a human-in-the-loop approach.

Personalized Recommendations

US-based DTC e-commerce company

Problem

Our client wanted to provide personalized recommendations for their DTC products (skincare). They were dealing with a cold-start problem (not enough interactions data to recommend a product to a user) as their product offerings change weekly.

Solution

We built a copilot tool as an AI mixture of experts to analyze the e-commerce site and offer actionable insights for conversion rate optimization. The copilot optimizes both layout, and which product to recommend to which user.

Outcomes

We achieved a 15% lift in RPV (revenue per visitor) vs a control group (previous static list)

Client references available upon request.