Grounded AI Assistant & Booking Agent
The assistant in the corner of this site is our own agent: it answers only from our live company knowledge, declines anything outside its remit, and books a meeting on a real calendar without a human in the loop.
2025/26
Ongoing
Xpergia
The Challenge
Most visitors to a consultancy website arrive with a specific question – have you done work like mine, what does this cost, can you help with X – and leave without asking it. A contact form answers nothing, and an email thread takes days to reach the point where anyone can tell whether there is a fit. We wanted to answer those questions at the moment they were asked, at any hour, and let a genuinely interested visitor book time with us there and then. The hard part was not building something that talks. It was building something we would be comfortable putting our own name behind: an assistant that never invents a capability we don't have, never wanders off into general chat, and never becomes a way to extract answers from us we didn't intend to give.
Our Approach
- Built the assistant as an agent rather than a chatbot – it retrieves, decides, and acts, with booking as a real tool call against a live calendar
- Grounded every answer in our own company content, so the assistant can only speak from what we have actually published
- Set a relevance floor so that when nothing relevant is retrieved, the assistant declines instead of guessing – the single most important design decision in the build
- Generated the knowledge base automatically from the website's own content, so the assistant is never out of date with what the site says
- Scoped the agent's tools tightly, with rate limiting and abuse protection in front of them, on the principle that an agent should be able to do exactly its job and nothing else
The Process
Define the remit
We wrote down precisely what the assistant is for – our services, our products, our case studies, how we work, and how to get in touch – and, just as importantly, everything it must refuse. That list became the specification for the guardrails rather than an afterthought.
Ground it in real content
The assistant retrieves from a knowledge base built from our published content and matches it against the visitor's question. Nothing reaches an answer unless it clears a relevance threshold, so the assistant's confidence is tied to whether it actually found something.
Give it one real action
Rather than a long list of capabilities, the agent got one action that matters: booking a meeting. It checks genuine availability and writes to a real calendar, which turns an interested visitor into a confirmed conversation without anyone in the middle.
Harden it
Rate limiting per visitor and per session, abuse protection at the edge, and a hardened prompt that resists attempts to redirect the assistant or talk it out of its own rules. We tested the failure modes as deliberately as the happy path.
Keep it current
Publishing new content to the website updates the assistant's knowledge automatically, with no separate release. A new case study becomes answerable within minutes of going live.
The Solution
A production AI agent that a prospect can use before they ever speak to us. It answers questions about our work from our own published content, with every claim anchored in something we have actually said. When it doesn't know, it says so and points to a person – which turns out to be the behaviour visitors trust most. When someone wants to talk, it books the meeting itself against a live calendar rather than dropping them back into a form. And it stays inside its remit: ask it to write a poem, help with your homework, or explain its own internals, and it politely declines and steers back to what it is for. It is the clearest demonstration we have of how we think agents should be built – grounded, narrow, guarded, and genuinely useful within those bounds.
Impact & Results
24/7
Answers Without a Human
100%
Answers Grounded in Our Content
Zero
Touches to Book a Meeting
Minutes
From Publishing to Answerable
Visitors get specific answers about our work at the moment they think to ask
Interested visitors book a meeting inside the conversation, on a real calendar
The assistant declines anything outside its remit rather than improvising
New case studies and articles become answerable within minutes of publishing
No separate content pipeline to maintain – the website is the source of truth
The same pattern underpins the grounded agents we build for clients
Explore other Case Studies
AI-Powered Learning Platform for Higher Education
A personalized Socratic tutoring agent for college students at scale – it coaches rather than answers, adapts to each student's mastery level in real time, and works across question types through a plugin architecture, on a multi-LLM foundation.
Scalable to 2M+ students, multi-LLM support, real-time mastery tracking
AI/ML-Driven Football Insights Platform
A real-time football analytics platform built from scratch on Microsoft Azure – an autonomous, event-driven agentic pipeline that reacts to live match events, runs inference, writes the narrative, and grades its own predictions against results with no analyst in the loop.
Real-time in-match insights in <5s, 92% prediction accuracy, pre-game reports in minutes
Cloud Migration & Application Modernization for a Scientific Publisher
Migrating an acquired on-premise monolith to AWS microservices, re-architecting a data science module to absorb a costly Solr upgrade, and embedding with the operations team to drive adoption of an entirely new way of working.
50% faster DS execution, monolith to microservices, jQuery to React
Data Pipeline & Visualization Platform for a Scientific Publisher
Unifying fragmented data from PostgreSQL databases, flat files, and internal systems into Snowflake via AWS Glue, then building Tableau dashboards that gave editorial, production, and commercial teams a single source of truth for the first time.
5 unified data sources, 60% faster reporting, real-time Tableau dashboards
Technology Stack
We've partnered with organizations across sectors, bringing industry-specific expertise to every engagement.
Python 
AWS 
Docker 
xpergia
intelligence, integrated.
From strategy to deployment, your AI transformation begins here
© 2018 Xpergia. All rights reserved.
info@xpergia.com 
Oxford, UK
HQ
· Kochi, Kerala, India