VoiceCart AI : A Voice-first shopping agent

Online shopping still makes you do the work: filters, tabs, comparing products by hand. About 70% of carts are abandoned, and 2.2 billion people live with a vision impairment that makes a screen full of small controls hard to use. In a real shop you would just ask someone. VoiceCart AI is that someone. You talk to the store and the store talks back, while the page changes as you speak. Say "an umbrella under 40 dollars, sorted by price" and the grid filters itself. Say "compare the second and third" and a comparison table opens with the best values marked. Ask "what do people complain about?" and it reads the cons from real review data. Then add items, change quantities, check out and place the order, all by voice. You can interrupt at any moment and it stops mid-sentence. Under the hood, the browser connects directly to the AssemblyAI Voice Agent API over one WebSocket. It streams microphone audio in and receives speech, transcripts and tool calls back. Our FastAPI server only issues a one-time token, so the API key never reaches the browser. Every store action is one of eight JSON-Schema tools (search, show, compare, update compare, update cart, show cart, checkout, place order), and each tool call visibly changes the page. The agent is grounded: it only states prices, specs and reviews that a tool returned. When data doesn't exist, like a 69th result or a discount code, it says so instead of guessing. Orders are placed only after a clear "yes". To make the voice feel natural, we built a playback queue in an AudioWorklet that cut audible gaps from 17–52 to 2–3 per reply, and a two-step interruption check that ignores the agent's own echo. The shopper hears an acknowledgement within 1.9 seconds, and 56 automated tests keep the tools honest. The agent runs in the browser and needs only a product API and a token endpoint, so any online store could add it to its existing site.