Ask Andi
Hosted on the client's site, not one I control.
- Year
- 2026
- Role
- Front-end Developer
- Stack
- GCP
- Vertex AI
- React
- Next.js
- TypeScript
- JavaScript
The problem
Prospective clients browsing AND Digital's website had no quick way to ask about the consultancy's capabilities or get connected with the right expert.
The approach
The client was already on GCP, so Vertex AI was the starting point rather than a choice. The harder problem was matching: prospective clients describe what they need in their own words, and early versions confidently routed those conversational asks to the wrong case studies and the wrong experts. The fix was to give the model a fixed set to match into (a dataset of expert fields from across the company), and then refine the prompts against real phrasing until the matches held up.
The guardrails
An LLM endpoint on a public marketing site is an open invoice, so abuse controls were a launch requirement rather than a hardening pass. Rate limiting and a ceiling on reply length stop a single visitor running up unbounded cost, and PII is stripped from messages before they reach any model. That last one came out of a deliberate security pass rather than an incident, which is the order you want it in.
The outcome
Live on AND Digital's website and actively routing enquiries to the relevant experts.