The brief
Every agency selling AI assistants should be running one, and most are not. We wanted the proof to be checkable in one click rather than described in a case study — so the assistant that answers questions about our services is the same kind of build we sell, and you can interrogate it while you read this.
What we built
Grounded in the same data as the website
The system prompt is generated at request time from the files this site is built on: every service and its features, the full FAQ, both rate cards, the delivery process and the markets we serve. The assistant and the pages cannot contradict each other, because there is one source and no second copy to update.
No vector database — on purpose
RAG exists because a corpus is too large to pass to a model in full. Ours is not. Retrieval infrastructure here would add cost, latency and a failure mode to solve a problem we do not have. The honest architecture for a small, well-structured corpus is to pass it in full — and to say so, rather than sell a vector database by default.
Streaming, so the first words arrive immediately
Answers stream token by token on Claude rather than appearing after a four-second pause. On a sales widget that difference decides whether a visitor waits for the answer.
A fallback that keeps it alive
If the model API key is missing or the stream fails mid-answer, it falls back to a local retrieval answer from the same knowledge rather than leaving a dead bubble on the screen. A widget that fails visibly costs more trust than one that never existed.
Enquiries captured, not just conversations held
Its own endpoint writes qualified enquiries out of the conversation, so an interested visitor at midnight becomes a lead rather than a transcript nobody reads.
Stack
- Claude
- Next.js
- TypeScript
- Streaming responses
- Local retrieval fallback
Where it stands today
It is live on every page of this site, including this one. Ask it what a RAG system costs, which of the fourteen services fits a problem you describe, or how the timezone overlap works for the US — then judge the build on the answer rather than on our description of it.
Services behind this build
Custom AI Assistants
A private AI assistant trained on your business — your products, your pricing, your process — not a generic chatbot.
See the serviceAI Chatbots & WhatsApp Bots
24×7 sales and support that qualifies leads, answers questions and books meetings while you sleep.
See the serviceRAG & Knowledge Systems
Turn your PDFs, contracts, SOPs and spreadsheets into an answer engine your team can just ask.
See the service
