Tenzr Oracle

Purpose

The aim was to make asking a question feel as simple as using a chat application. The complexity belonged behind the interface: understanding which data a question referred to, composing a query, and presenting an answer or chart.

This was an internal tool in a healthcare environment, so data access and privacy requirements were part of the design from the beginning.

Build or buy

I researched existing software, integration options and the risk of becoming tied to one vendor. I also worked with colleagues on a company data-warehouse initiative. That work helped frame the problem beyond connecting an LLM directly to individual tools.

Choosing the capabilities we needed, and deciding where existing software could help, took substantial research alongside the implementation.

Queries and controls

I implemented Claude tool use and a semantic layer that composed SQL from defined models. The implementation included read-only query enforcement, streaming responses, rate limits, token tracking and cost controls.

A simplified view of the query flow.

Read-only queries limited the kinds of operations the tool could perform. Model-usage controls addressed a different concern: how requests consumed tokens and incurred cost. Both were part of the work behind the chat interface.

Architecture changes

In an earlier version, I built a TypeScript/Next.js web service and a Python/FastAPI graph engine, deployed independently on Amazon EKS. GitHub Actions built each service separately and pinned releases to a commit.

The graph engine was removed during the August 2026 rebuild around Bedrock. That earlier architecture is part of the project's history; it is not a description of its current deployment.

Delivery

A working internal application that connected a plain-English question to defined company data models, SQL, and an answer the person asking could use. My contribution spanned the interface, backend implementation and deployment, alongside research into the wider set of tools and integrations.