Analytics Agent Platform

What it is

A multi-agent analytics platform built for a major entertainment brand's internal teams. It lets non-technical executives ask questions in plain language — about talent, audiences, partnerships, revenue and viewing — and get back grounded, data-backed answers instead of navigating a wall of dashboards. I currently build and maintain it.

What I work on

• Maintaining and extending the live platform that the client's teams use day to day.
• Led the authentication migration to an enterprise identity setup with single sign-on — designed as a reversible cutover that didn't disrupt existing users.
• Built the tooling that evaluates agent responses and helps debug real chat issues coming out of production.
• Greenfielded a dedicated ROI agent on the same platform — natural-language questions about partnership revenue, hours-of-viewing and series ROI, answered from governed warehouse data.

The agent harness

The platform runs on a custom, configurable agent harness that's assembled per request rather than hard-coded. Each agent is composed from four primitives:

• System messages — the base instructions that define an agent's behaviour.
• Skills — focused, trigger-gated prompt modules that load only when a question calls for them.
• Resources — structured data and runtime configuration (current date, fiscal period, defaults) materialized at request time.
• Tools — grouped, in-code implementations the agent calls to query the underlying data.

Because the configuration lives in data rather than code, agents can be reshaped — new skills, different tool groups, a different model — without a redeploy.

Keeping answers trustworthy

An analytics agent is only useful if executives can trust the numbers, so evaluation is a first-class part of the work:

• promptfoo-driven evaluations gating agent behaviour as prompts and tools change.
• Conversation-replay regression tests captured from real user feedback and scored with an LLM-as-judge against expected behaviour and tool usage.
• Full test suites covering the tooling and the API surface end to end.
• Chat-debugging tooling to trace a single conversation through prompt assembly, tool calls and results when something looks off.