Calendar apps are important for productiveness however it’s laborious to distinguish sufficient to have sustained development from simply the core utilization. Y Combinator-backed Superpowered, which is an AI-powered notetaker on your conferences that doesn’t contain recording bots, hit this roadblock and is now pivoting to change into Vapi, an API supplier so anybody can simply create a natural-sounding voice-based AI-powered assistant.
Superpowered was based in 2020 by Jordan Dearsley and Nikhil Gupta. However after three years of engaged on it, Dearsley stated the workforce needed to work on the tougher product. The corporate isn’t shutting down the preliminary product because the startup stated that Superpowered is worthwhile — it’s within the strategy of bringing somebody in to run it. Y Combinator stated in June that greater than 10,000 individuals had been utilizing the product weekly, however the firm didn’t present any up to date numbers.
Thus far, Superpowered/Vapi has raised $2.1 in seed cash from traders together with Kleiner Perkins and Summary Ventures.
Pivot to Vapi
The corporate gives Vapi as an API to let builders create a bot utilizing simply prompts — it then put it behind a telephone quantity. Moreover, it gives an SDK integration so builders can embed the bot on web sites and cell apps.
Dearsley advised TechCrunch over e mail that the concept to construct Vapi stemmed from a private downside. He had moved to San Fransisco and began lacking his family and friends, who had been in a unique time zone. He constructed an AI bot connected to a telephone quantity on the opposite finish to speak to somebody in an effort to kind his ideas.
“I preferred it, however I used to be regularly annoyed with how unnatural it was. It wasn’t like speaking to an individual. The voice sounded off, there could be lengthy delays earlier than it responded, and it will interrupt me whereas I used to be talking.” he stated.
“So I stored engaged on it and going for my walks with it. Finally, we received fascinated with this dialog downside. It’s actually laborious to make one thing really feel human. Voice assistants right now are clunky and turn-based, we wish to construct one thing that feels human.”
Technically, Vapi is at the moment stringing a bunch of third-party APIs to construct a sturdy voice dialog platform. As an illustration, it makes use of options from Twilio for telephony, Deepgram for transcription, Daily for audio streaming, OpenAI for responses, and PlayHT for text-to-speech.
ScaleConvo, a startup within the YC winter batch for 2024, is already utilizing Vapi to launch conversational bots for gross sales groups and property administration firms. Nonetheless, Vapi didn’t disclose its different purchasers. The corporate is opening up its API with Vapi Telephone and Vapi Net merchandise right now.
Challenges for Vapi
One of many largest challenges the startup has is to scale back latency, based on Magnus Revan, an ex-Gartner analyst and chief product officer at multimodal dialog startup Openstream.ai.
“OpenAI fashions want between 2-10 seconds to generate a solution – whereas on the telephone the gold commonplace is to have 700ms between the consumer ending speaking after which the ‘bot’ beginning to discuss. And attending to sub 1-second latency with succesful fashions (excessive parameter depend open-source fashions like LLaMA2 70B) is actually laborious,” Revan stated.
Presently, Vapi has a latency of 1.2-2 seconds relying on varied elements. Dearsley expects to deliver down latency to below one second within the subsequent month because of Vapi’s personal work and OpenAI’s enhancements.
Mohamed Musbah, an angel investor in Vapi additionally stated that the startup’s answer will enhance with general advances in API.
“As OpenAI and others enhance their fashions, Vapi’s platform will change into extra highly effective, outfitted with higher data bases, code execution capabilities, and bigger context home windows. Vapi’s concentrate on fixing the best friction areas in voice communication shall be its edge as consumer demand grows for voice assistants,” he stated.
Nonetheless, this places the onus on the development of different options somewhat than Vapi itself. Dearsley stated that reliance on different APIs reduces Vapi’s defensibility if huge firms begin transferring into that space. Nonetheless, the workforce stated that it has an edge when it comes to having constructed infrastructure to deal with 1000’s of calls concurrently. Dearsley emphasised that with Vapi’s net and telephone API launch for the general public, the workforce can even look to construct its personal fashions for audio-to-audio options.