Tandem

Inspiration

Successful livestreams often rely on multiple hosts: one presents products while another manages the operational workload; displaying products, creating promotions, and responding to chat. Smaller businesses, however, frequently have only a single host juggling all of these responsibilities, resulting in missed customer questions, delayed promotions, and reduced engagement.

Our solution is an AI-powered co-host that automates these repetitive tasks, allowing sellers to focus entirely on engaging with their audience.

How it works

Our system consists of four specialized AI agents working together throughout a livestream.

🎙️ 1. Audio Parser

The Audio Parser continuously listens to the livestream and understands what the host is talking about.

OpenAI Whisper-1 generates a real-time transcript from the livestream audio. The transcript is passed to GPT-4o mini, which extracts the host's intent. When a product is mentioned, the agent identifies the corresponding product from the product database and automatically displays it on the frontend. If the host announces a promotion (e.g. "10% off for the first 40 customers"), the agent automatically generates and surfaces the appropriate discount coupon.

This removes the need for a second operator to manually switch products or create promotions during the stream.

💬 2. Chat Classifier

Livestream chats move quickly, but not every message requires a response.

Using GPT-4o mini, incoming chat messages are classified into two categories:

No response required "You're so pretty!" Requires host attention "Can this bag fit an iPad Pro?"

Only messages that genuinely require assistance are forwarded to the next agent, significantly reducing noise for the host.

🤖 3. AI Responder

The Responder generates context-aware replies for customer questions.

Powered by GPT-5.4 mini, it retrieves information from multiple sources, including:

the product currently being showcased, the user's chat message, the product database.

Examples:

Current product: Rhode Lip Tint

"How much is this lip tint?"

The agent retrieves the product information and responds:

"Hi! This lip tint is $6.40."

If the question is unrelated to the current product:

"Can this bag fit an iPad Pro?"

The agent searches the product catalog for relevant bag products before responding:

"Which bag are you referring to?" or "Yes! Our Croissant Bag comfortably fits an iPad Pro and can also fit a MacBook."

This allows the AI to answer questions accurately even when customers ask about products outside the current demonstration.

📊 4. Post-Stream Debrief (Planned)

Due to time constraints, this feature was not fully implemented.

After each livestream, the AI would generate an actionable summary including:

Products ranked by customer demand based on chat activity. Unanswered customer questions that require follow-up. Customers who expressed purchase intent but did not complete a purchase, helping sellers identify potential leads for re-engagement.

Why we built it

Rather than replacing livestream hosts, our AI acts as a virtual co-host, automating repetitive operational tasks so sellers can spend more time interacting with customers and driving sales. The result is a smoother livestream experience, faster customer support, and better conversion opportunities; especially for small businesses running streams with limited manpower.

Built With

• amazon-web-services
• lambda
• react