SYS.01 Objective
Multifi wanted to provide fast, reliable AI-powered finance answers directly through their website, without relying on third-party tools, and without compromising on brand tone or compliance.
That combination rules out most of the market. A bolt-on chat widget solves the interface problem and creates three others: customer questions are routed through somebody else's platform, the assistant's voice is configured rather than written, and the operator inherits a second vendor relationship for something that is, technically, one API call.
The brief was therefore not “add a chatbot”. It was: own the endpoint.
SYS.02 Creative response
A custom WordPress plugin that creates a secure, REST-based chatbot endpoint connecting directly to OpenAI's GPT-3.5-Turbo API. The site already ran WordPress, so the plugin sits where the traffic already is: no new platform, no new login, no new subdomain.
- Chat endpoint. A single REST route handling real-time responses. The browser talks to multifi's own domain; the plugin talks to OpenAI server-side, which is what keeps the API key off the front end.
- Rate limiting and sanitisation. Both built in, both applied at the endpoint before a request is allowed to reach the model, the two safeguards that stop a public chat route becoming a public billing route.
- Custom system prompts. The assistant's voice is written, not selected from a dropdown, with prompts constructed to reflect multifi's tone of voice.
- Admin settings. API key input and validation from inside WordPress admin, so key rotation is a task for whoever runs the site rather than whoever wrote the plugin.
- Testable shortcode. The chat interface can be dropped into any page or post to exercise the live endpoint, a staging surface as much as a delivery mechanism.
- Mixpanel tracking. User analytics on the conversation, so the behaviour of the thing can be observed rather than assumed.
Design decisions
Three choices define the build. Server-side by default; the model is never called from the client, so the credential never leaves the host. Safeguards before intelligence, rate limiting and sanitisation are cheap at build time and expensive to retrofit after a bad week. And the plugin is the product: everything needed to run, key, test and observe the assistant sits inside one installable artefact.
It is deliberately lightweight. No vector store, no retrieval layer, no fine-tune; the behaviour comes from a well-written system prompt against a general model, which is an honest description of the scope rather than a limitation being hidden.
SYS.03 Design decisions
SYS.04 Results
GPT-3.5-TURBO, OpenAI model called behind the endpoint
SHORTCODE + ADMIN, testable shortcode, API-key entry and validation in WP admin
What the brochure claims for it is a capability claim, and is repeated here as one: the plugin let multifi embed a branded, consistent chatbot on their site, supporting customer engagement without additional tools or platforms, with usage safeguards and easy API key management, while retaining full control over data and messaging.
This is a capability build. No conversation volume, deflection rate, containment rate, answer-accuracy score, uptime figure, latency measurement or commercial outcome has been published for it, and none is claimed here.
Want the endpoint, not the platform?
Secure, rate-limited, prompt-controlled AI endpoints built into the stack you already run, and documented plainly enough to hand over.
If this looks close to something you are dealing with, the quickest route is to say so in a sentence. You will get a real reply from a person, usually with a first read on the problem rather than a calendar link.