Case study · Built and run by Foldox
hiy.ai
AI twins and support agents that answer from your own content.
- Kind
- AI twins and support agents
- Platforms
- Web app and embeddable widget
Illustration with sample data.
The problem
Experts and support teams answer the same questions again and again. An assistant can help, but only if it answers from their own material, shows where each answer came from and knows when to hand over to a person.
What we built
- An AI twin that learns from websites, YouTube transcripts and documents.
- Answers in the owner’s voice, with the sources behind each answer.
- A chat widget and an inline embed that work on other websites.
- Labelled as AI on every surface, including embeds on other websites.
- A report link on every public page, feeding a review queue.
- Account deletion and a full data export in the dashboard.
Approach
Key decisions. What we chose, and why.
- 01
Sources for answers, not for training
Sources are used to find answers. They don’t train a shared model.
- 02
Several model providers
hiy.ai works with models from Anthropic, OpenAI, Google and DeepSeek, so it isn’t tied to one provider.
- 03
A safety floor that can’t be switched off
Messages about self-harm or minors are caught before any model sees them, and handed to a person.
- 04
Honest about being an AI
The AI label can’t be removed on any plan, and the agent says it is an AI when asked.
Stack
What it runs on.
Web
- Next.js
- Vercel
AI
- Anthropic Claude
- OpenAI GPT
- Google Gemini
- DeepSeek
- Vercel AI Gateway
Data and email
- Supabase
- Postgres
- Resend
Names are trademarks of their owners.
Capabilities shown
What this work shows. Each linked tag is a service we offer.
- Chat over your data
- AI agents
- AI app development
- Citations
- Embeddable agents
- Multi-model
Start a project
Want an assistant that answers from your content?
Tell us what it should know, and who will ask. We aim to reply within two working days.