Sale Helper

• 

  Testing out the whisper live function

• 

  Fine tuning the prompts!

• 

  Finally got the code to run properly

Inspiration

"To help new salespeople effectively address customer queries and provides relevant technical information through a sales helper."

Many new salespeople struggle with seamlessly guiding conversations to resolve customer queries and drive towards closing deals. This challenge often stems from a lack of experience and confidence in handling diverse customer concerns. As a result, sales interactions can become unfocused, leading to missed opportunities and customer dissatisfaction.

This project is designed to assist sales teams by leveraging customer relationship management (CRM) systems to streamline their daily workflows throughout the end-to-end sales processes. By integrating artificial intelligence (AI) models, the project aims to provide smart sales helper functionalities that enable managerial insights and optimize sales strategies.

What it does

As the conversation progresses within the call, insights on how to provide context to the customer's queries are generated to aid the salesperson. These sales insights provide contextual knowledge that triggers keywords or directions that steer the conversation to clarify a customer's doubts. Instead of needing to scan through an entire FAQ or memorize the script, the user can use the information provided in the 'bot insights' segment to provide that information to the customer.

How we built it

1️⃣ Faster-Whisper: WhisperLive
We utilized the WhisperLive library in order to implement the ASR component of our solution.

2️⃣ Initial Filtering Process
A filter was created using openAI to see if the transcriptions coming in were worth of query to the vector database. If transcripts passed the filter, then the transcriptions would be refined further into a more concise query, again using the openAI model.

3️⃣ Retrieval Augmentation Generation (RAG)

• We web scraped a few raw html pages from the Tiktok Business Center using BeautifulSoup4 and then used OpenAI to restructure the page into text format.
• We vectorized each text using the openAI model and then loaded the embeddings into a vector database. We selected Qdrant to be our vector database for this project.
• Once the vectorized pages were loaded, we added the retrieval into our pipeline to fetch the nearest vector result, and used OpenAI to restructure the prompt based on the live transcriptions that were coming in through the ASR.
• To ensure queries were complete, we used openAI to complete or refine any incomplete or incoherent sentences to form a short query.
• Once a transcription passes the filtering process, the query is vectorized and then we fetch the nearest result. The result returned is then provided as context to OpenAI to then craft a relevant technical tip that could help the salesperson.

4️⃣ Saleshelper Bot Insight The generation is then pushed to the front end to aid the client-facing user as an insight.

Accomplishments that we're proud of

Having a working product that can actually add value!

What we learned

Everyone came from a different background and had a different work style. By focusing on our strengths and pooling together our resources, while understanding our differences, we can accomplish what none of us can do alone.

What's next for Sales Helper

Making this relevant for many more domains!

Built With

• axios
• getstream
• openai
• python
• python.socketio
• react
• socket.io
• tailwind