The work, on the record.

Production AI before the playbook existed

The engagement started where most wouldn't dare: the GPT-3.5 era, when Retrieval-Augmented Generation was more theory than practice. The sales academy had a five-day problem — that's what it took to hand-author a single training nugget — and one hard constraint: every generated artifact had to come exclusively from the academy's own curated content. In regulated insurance, a hallucinated product detail isn't a bug. It's a liability.

Pangea made the case for grounded generation, and my CTO led the architecture. Months of failures and iterations followed — working out RAG patterns that weren't written down anywhere yet. The system that emerged produced grounded training videos in under fifteen minutes, with image-library integrations cutting the manual asset hunt out of the pipeline. The build adopted MongoDB's vector database early enough that MongoDB's own product team invited the architects in to compare notes.

The same grounded foundation powered a second surface: a live co-pilot that coached agents through the moments that kill deals — "I can't afford it," "there are other providers," "I need to check with my family" — turning hesitation points into handled conversations.

Around the person, not the model

Training content was the opening move. The full workflow wrapped the agent's entire working day — grounded in the carrier's own products, process, customer types and objections:

Experience parity, then scale

The brief was reproducing what good care feels like in a clinic — through a screen, at the pace of a chain that compounded month over month. The team was built from the ground up, DevOps included, and the pet's full lifecycle went digital one capability at a time: online booking · payments and subscriptions · repeatable clinic onboarding · veterinary-management-system integrations · medical records that follow the pet across visits and locations.

The through-line was experience parity — whatever felt thoughtful and personal in the clinic had to feel the same online. Every system was designed to be onboarded onto, integrated with, and extended, because the network was growing while it was being built. One clinic became eighty-plus, from >90% phone-led to majority-digital booking.

One job, done autonomously

With the platform established, the next friction was booking itself — a few minutes and a few too many clicks. The answer was a fully autonomous agent with exactly one purpose: book appointments in natural language. Deliberately scoped to booking and nothing else, then hardened around that boundary. Built and ready for A/B testing in two weeks, rolled out to low-impact locations first.

The beta became an adversarial testing ground for free: early users poked at the agent — who won the Super Bowl, who took the NBA title, directions from A to B. Every off-script session surfaced an assumption and closed a guardrail gap. The agent learned to stay on purpose, decline everything else however creatively asked, and shut its doors on persistent abuse.

Live across all geographies, the most important number was the one that didn't move: traditional bookings held steady. The agent added net-new bookings on top — new user acquisition, zero cannibalization — and GoodVets had a genuine AI use case in production.

Three systems, one truth

Growth produced data everywhere — payments in one system, the care platform in a second, practice management in a third — and the important questions lived between them. A medallion-architecture warehouse (Bronze, Silver, Gold) on AWS pulled all three into one governed source of truth: raw stays replayable, standardization happens in one place, and business logic lives in a star schema that dashboards build on without touching the pipeline.

The contrarian call was building the pipeline custom instead of buying an ETL tool — made affordable by externalizing every source definition and mapping into plain configuration. That design let a three-person backend team with no prior data-engineering experience stand the whole warehouse up in under two months, on PySpark and Apache Iceberg. And because the base was open and owned, AI quality checks sit between the layers — flagging source-data changes before they break the pipeline, not after they break a dashboard.