Ellum

Inspiration

We've worked on e-commerce stores (Shopify/etc), and managing customer support calls was tedious and expensive (small businesses online typically have no number for contact), despite that ~60% of customers prefer calls for support questions and personalization. We wanted to automate the workflow to drive more sales and save working hours.

Existing solutions require custom setup in workflows for chatbots or asking; people still have to answer 20 percent of questions, and a lot are confirmation questions (IBM). People have question fatigue with bots to get to an actual human.

What it does

It's an embeddable javascript widget/number for any e-commerce store or online product catalog that lets customers call, text, or message on site chat about products personalized to them, processing returns, and general support. We plan to expand out of e-commerce after signing on 100 true users who love us. Boost sales while you are asleep instead of directing customers to a support ticket line.

We plan to pursue routes of revenue with:

• % of revenue from boosted products
• Monthly subscription
• Costs savings from reduced call center capacity requirements

How we built it

We used a HTML/CSS frontend connected to a backend of Twilio (phone call, transcription, and text-to-speech) and OpenAI APIs (LLMs, Vector DBQA customization).

Challenges we ran into

• Deprecated Python functionality for Twilio that we did not initially realize, eventually discovered this while browsing documentation and switched to JS
• Accidentally dumped our TreeHacks shirt into a pot of curry

Accomplishments that we're proud of

• Developed real-time transcription connected to a phone call, which we then streamed to a custom-trained model -- while maintaining conversational-level latency
• Somehow figured out a way to sleep
• Became addicted to Pocari Sweat

What we learned

We realized the difficulty of navigating documentation while traversing several different APIs. For example, real-time transcription was a huge challenge.

Moreover, we learned about embedding functions that allowed us to customize the LLM for our use case. This enabled us to provide a performance improvement to the existing model while also not adding much compute cost. During our time at TreeHacks, we became close with the Modal team as they were incredibly supportive of our efforts. We also greatly enjoyed leveraging OpenAI to provide this critical website support.

What's next for Ellum

We are releasing the service to close friends who have experienced these problems, particularly e-commerce distributors and beta-test the service with them. We know some Shopify owners who would be down to demo the service, and we hope to work closely with them to grow their businesses.

We would love to pursue our pain points even more for instantly providing support and setting it up. Valuable features, such as real-time chat, that can help connect us to more customers can be added in the future. We would also love to test out the service with brick-and-mortar stores, like Home Depot, Lowes, CVS, which also have a high need for customer support.

Slides: https://drive.google.com/file/d/1fLFWAgsi1PXRVi5upMt-ZFivomOBo37k/view?usp=sharing

Video Part 1: https://youtu.be/QH33acDpBj8 Video Part 2: https://youtu.be/gOafS4ZoDRQ

Built With

• html
• javascript
• llm
• machine-learning
• modal
• openai
• python
• twilio