Regular updates on the latest VC-backed AI startups. Follow along to stay informed!
We recently attended a demo day for Sandbox. Of the 16 student-led startups that pitched, 9 claimed to use AI in some way. This seems roughly representative of our anecdotal experience in the startup scene ever since ChatGPT became the fastest app ever to reach 100M users - more than half of new startups are focused on AI. In today’s special edition, we are covering some companies from the Sandbox demo day which have not yet raised institutional VC funding.
Replay
Problem to be Solved
Training new sales reps is difficult and expensive. If training is done poorly, reps churn before producing any revenue. Even if it’s done well, reps take months to ramp. Replay aims to reduce rep churn and shorten ramp cycles by using AI to create live role-play-based sales training.
How They Use AI
Replay is built using OpenAI’s GPT APIs (e.g., GPT-3, ChatGPT). It uses a series of prompts to get the model to respond in the voice of a certain type of potential customer with concerns selected by the user. In addition to the proprietary prompt engineering, Replay is developing a solution for giving the foundation model additional context from a database of recorded sales interactions.
Business Model
Replay sells access to its platform as a service on a per-rep subscription basis. The first version of the tool is getting early traction at a growing number of companies like Vivint, where founder Chase Meredith spent a summer selling door-to-door. Replay is actively raising its first round of investment.
VIVID Interiors
Problem to be Solved
Interior design is prohibitively expensive, so most people’s living space is uninspiring. VIVID Interiors uses AI to generate beautiful, customized, shoppable interior designs.
How They Use AI
VIVID Interiors is built using stability.ai’s Stable Diffusion, which is a text-to-image or image-to-image model. The user uploads an image of their room and selects from a range of design styles. Behind the scenes, VIVID calls Stable Diffusion with a low-strength version of the image and a prompt-engineered text description of the desired design style. An image of the redesigned room is then returned to the user. VIVID Interiors then uses an image recognition service to search the web for real furniture and decorations similar to what was generated in the photo.
Business Model
Vivid Interiors currently operates a low-cost version of the traditional interior designer business model, charging early users a flat fee for each room designed. In the future, founder Chalise Moore aims to add a commission-based revenue stream from the furniture and decorations sold through the platform.
Mindsmith
Problem to be Solved
Course design is a bottleneck in teaching and training. Research by an eLearning consulting firm shows that developing a new course can take from 20 to 200 hours. Mindsmith founder Ethan Webb claims his AI-powered lesson generator shortens this timeline to less than one hour.
How They Use AI
Mindsmith uses OpenAI’s transformer-based LLM models (both GPT-3.5 and GPT-4) and unique prompt engineering to generate learning content. Additionally, Mindsmith gives the model context about the specific learning topic by using vector analysis to search embeddings of instructional material and data uploaded by the user.
Business Model
Classic SaaS. Mindsmith has quickly scaled via product-led growth to over 2,400 users including teams at Google and the University of San Francisco. The company received interest from investors on demo day and is in the process of raising a $500K angel round.
Buster
Problem to be Solved
Business people who don’t know SQL can’t make data-informed decisions without bugging the data team, which is always backlogged. Buster thinks AI can enable these business people to answer their own data questions using natural language.
How They Use AI
Buster uses OpenAI’s models to generate SQL queries in response to natural language requests. This is a non-trivial problem that many have set out to solve. The challenges of text-to-SQL mainly arise from providing the correct context to the model in order for it to write the correct query (including joins) based on the actual tables and columns in the database (which can be very large). Additionally, Buster likely implements self-healing code in order to correct queries that are only marginally wrong.
Business Model
The usual - SaaS with starter/team/org tiered pricing. Buster is offering users of the early product (which can currently query 15 tables at 98% accuracy) beta pricing for life.