Production-grade RAG for an
insurance sales academy

An insurance provider's sales academy had a multi-fold problem: keeping sales agents current on a constantly evolving product line, and on the protocols for handling a customer call. At the time, all of it ran on human effort — the academy team manually authored every training nugget and published the scripts agents leaned on during live interactions.

Two hard constraints shaped everything. Every artifact had to be generated exclusively from the academy's own curated content, and it all had to work reliably enough to put in front of live agents. I came on as Lead Architect and CTO. This was the early era of large language models — we were still on GPT-3.5, and Retrieval Augmented Generation was more theory than practice. But we could see the solution write itself. We made our case, convinced the client this was the right direction, and set out to build something none of us had built before.

---

The challenge

The academy didn't just want AI-generated content — they wanted AI-generated content that could only ever draw from their approved material. In a regulated insurance context, a model that hallucinated a product detail or invented a claim wasn't a bug, it was a liability. That made grounding non-negotiable. Layered on top were the realities of building on frontier technology before the playbook existed:

• No established RAG patterns. The concepts we needed weren't written down yet — we had to grasp them through experimentation, failure, and iteration.
• A hard content boundary. Every generated artifact had to be traceable back to the academy's own curated database, with nothing invented.
• A painfully slow status quo. Producing a single training nugget took the academy team a good five days of manual work.
• Two very different surfaces. The same grounded intelligence had to power both long-form training content and fast, in-the-moment answers on a live call.

---

The approach

We committed to Retrieval Augmented Generation as the backbone and built a full-stack solution around it, grounding every output in the client's curated content so nothing was ever generated from outside their approved material. Over a few months of numerous failures and iterations, we worked out the unwritten concepts and the nitty-gritty that separated a demo from something production-grade.

Along the way we were early adopters of MongoDB's vector database capabilities — early enough that we were invited to sit down with MongoDB's product team to share what was working and what wasn't. That gave us both a front-row seat to a maturing technology and real influence over it.

The result was a solution that did two very different jobs from the same grounded foundation:

In a regulated insurance context, a model that hallucinated a product detail wasn't a bug — it was a liability. Grounding was non-negotiable.

The same grounded foundation powered both surfaces: long-form training content that took minutes instead of days, and fast, in-the-moment answers an agent could trust mid-call. Neither was allowed to step outside the academy's approved material.

---

Outcomes

An experimental, long-shot idea turned into a real, working capability for the sales academy. More than the individual features, the project proved a point that was far from obvious at the time: in the GPT-3.5 era, production-grade AI wasn't a myth. We showed a real team, in a regulated industry, that a grounded RAG system could be trusted with live, business-critical work.

---

Tech & tools

---