Trusted AI chatbot for 20,000+ dental professionals

Client

The International Team for Implantology (ITI) is the leading global association in implant dentistry, with more than 20,000 members in over 100 countries. Its aim is to advance implant dentistry for the benefit of the patient, through education and research. They came to us to make that knowledge easy to reach through a single conversational channel, giving members instant access to clinical answers.

Challenge

• Siloed knowledge base: ITI's evidence-based content had built up across many separate platforms and formats over decades, from a book series and journals to a clinical guidelines database and an online academy. Getting to the right answer meant knowing which resource to open, and that was not always quick.
• Content consumption: Members increasingly expected to ask a question and get a direct answer - not open three tabs and hunt.
• Retention: ITI's engagement with members was strongest at conferences, events, and study club meetings. The organization wanted to strengthen its everyday presence, putting its knowledge within reach at any moment. That constant availability also gave the community an additional reason to renew their membership.
  

Solution

Scoping to a safe first step

ITI came to us with a broad, multi-year concierge vision and a backend-only proof of concept. We narrowed it to a smaller, deliverable first project: an educational AI Concierge grounded only in ITI's own treatment guides and clinical guidelines. We scoped it as an educational product, not a diagnostic or treatment tool. We started from scratch: a complete new PoC, then an MVP built on top of it.

Today the chatbot runs as a standalone web app. ITI plans to build it into their new mobile app for members, where it is set to be one of the main features.

Quality from data preparation, not a bigger model

The chatbot's accuracy does not come from the size of the model. It runs on a deliberately small one, GPT-5-mini with low reasoning; the quality comes from how carefully we prepared the data behind it. That meant cleaning mixed-format guides and publications into consistent text, a tested chunking strategy, embeddings, retrieval modules built per content type, and a system prompt engineered to prevent the model from going beyond the source material.

With this preparation the model does not invent anything; it just rewrites ITI's vetted content into a clear, sourced answer. That keeps a tiny model highly accurate and free of hallucinations, while holding cost and response times low.

An engineering system that proves it

Because no one could hand-check clinical answers at the volume we were generating, quality had to be measurable, and we tested it in layers rather than one pass. With ITI we wrote a versioned Evaluation Framework defining what a good answer is, then built that definition into an automated suite using Promptfoo, with every run tracked in Langfuse:

• A dataset of expert-written questions, each tied to its expected source document, so we test retrieval as well as the answer, plus dedicated sets for each rule the chatbot must keep: stay on topic, name no brand, give no direct medical instructions, admit when it does not know, and never invent facts.
• Scoring on every meaningful change for metrics like Context Faithfulness, our main check that an answer stays true to its sources, with results visible in production so any drop can be traced to its cause.
• Load testing before the Istanbul launch, which caught a rate limit and a database bottleneck in time to fix both.
• Hands-on exploratory and safety testing, with people deliberately probing for unsafe and adversarial inputs.

Pragmatic build choices

We started from Chatguru, Netguru's own open-source chatbot platform, which gave us a working foundation and saved about a month of groundwork. The whole system runs on Azure inside ITI's own environment, so member data never leaves their infrastructure.

To pick the right retrieval architecture, we ran two approaches side by side: a standard SQLite vector store and a graph-based LightRAG. After comparing both, SQLite came out ahead: faster, cheaper, and sufficient for the questions practitioners actually ask. LightRAG stays in the codebase as a foundation for a future advanced mode.

The build moved through three clear phases: a fixed-scope discovery, a one-month proof of concept, and an MVP delivered in five sprints. It launched to several hundred practitioners at ITI's Istanbul conference in May 2026.

QR code + single sign-on: no account, no barrier

Members log in through a QR code and single sign-on (components built by other team), with no separate account to create. They can show the code at a study club or pass a photo around a conference table, and colleagues can scan it and start asking questions on the spot. This low barrier turns the chatbot into something members naturally share with one another.

Results

• 98% answer accuracy
• Near-zero production hallucinations
• 2s / 10s to first token and to a full answer
• 9/10 NPS, scored independently by both client stakeholders
• 140 documents ingested into a live, membership-gated web app
• 24 user tests across 12 countries, before the launch
• 100% of early survey respondents (15 so far) reported no answer they felt was factually wrong or unsafe