LABS

Labs

Working demos showcasing how AI can turn customer-experience data into decisions. Each one is a real, functioning demo — trained on a synthetic dataset.

RAG OVER VOICE OF CUSTOMER

The Listening Post

Why it exists

Customer feedback is scattered across reviews, surveys, and win/loss and churn notes, so teams fall back on hunches. The Listening Post lets you ask questions of that feedback in plain language and get answers grounded in the actual customer feedback (verbatims), with every claim cited back to its source.

What it does

Ask a natural-language question, optionally scoped to a segment, and it retrieves the most relevant customer feedback, passes them to the model as context, and streams back a cited answer. Follow-up questions keep the conversation in context.

How it's built

Rails 8 · PostgreSQL + Voyage AI embeddings · hybrid retrieval (semantic search + lexical search) · Claude API, streamed token-by-token over Server-Sent Events · deployed on Heroku.

Open the lab

MCP SERVER FOR VOICE OF CUSTOMER

Signal Line

Why it exists

Customer intelligence usually lives in a separate dashboard from where the work actually happens. Someone has to remember to go look. Signal Line puts this dataset's tools and data behind the Model Context Protocol, so an AI assistant like Claude can answer questions and take action on it directly, inside the tools people already use.

What it does

Search feedback, pull theme trends and segment health, and flag an account for review, live, over a documented MCP connection. It also ships a few ready-made prompts for common CS workflows, like a weekly at-risk brief and a per-account deep dive. The catalogue on the demo page reflects the running server's real tools, resources and prompts, so it's never out of date with what's actually callable.

How it's built

Rails 8 · StreamableHTTPTransport · bearer-token auth · tool/resource/prompt catalogue · deployed on Heroku.

Open the lab

AUTONOMOUS AGENT FOR C-SUITE BRIEFING

The Monday Brief

Why it exists

The most valuable signal in a customer base is usually the thing nobody thought to ask about — a health score sliding, a account flagged for review that nobody escalated, feedback data points that nobody connected to a pattern. The Monday Brief investigates on its own schedule and surfaces it to the CCO before anyone has to ask.

What it does

A scheduled agent run investigates at-risk accounts with its own tools: checks account health scores, recent feedback, usage trends, accounts flagged for review — then writes an executive risk brief naming the highest-risk accounts with data-backed reasons. The plan-act-observe trace is visible live as it works, and every run is kept in an archive.

How it's built

Rails 8 · Anthropic tool-use loop · Solid Queue background job · live-streamed output · genuinely scheduled Mon 7:02am AEST · deployed on Heroku.

Open the lab

VISUALLY EXPLORE CLUSTERED FEEDBACK

The Observatory

Why it exists

Customer feedback arrives as thousands of separate fragments — a review here, a survey response there — with no single view of how issues relate, grow or drift over time. The Observatory lays the whole corpus out as a map organised by meaning, so patterns that are invisible one verbatim at a time become visible at a glance.

What it does

Every piece of customer feedback is grouped into clusters and summarised, coloured by sentiment, theme or segment, sized by volume and positioned by how closely clusters relate in meaning. A time window filters the view to any range, and clicking a cluster or verbatim opens its full detail with a prefilled link to ask the Listening Post about it.

How it's built

Rails 8 · UMAP + HDBSCAN · Named Clusters · deployed on Heroku.

Open the lab

AGENT EVALS FOR SAFE DEPLOYMENT

The Proving Ground

Why it exists

Shipping a new agent prompt is a bet, the only way to know if it actually handles a hardship disclosure, a prompt injection attempt, or a customer who buries the real question in a ramble is to throw exactly those cases at it before a real customer does. The Proving Ground runs exactly those cases against every version of the agent, so a regression shows up in a verdict grid instead of in production.

What it does

Six simulated customers hold a real conversation with the support agent, each working from a secret brief it never sees. They improvise, and they decide for themselves when they've been helped or given up. A separate judge then scores the finished transcript pass or fail against six fixed criteria, it sees the briefs and every tool call, so it can catch an answer that sounded fine and missed the point. The result is a verdict grid for every agent version.

How it's built

Rails 8 · 6x Customer Personas · 4x tool-calling Support Agents · 1x Simulated Judge · deployed on Heroku.

Open the lab

SYNTHETIC VOICE OF CUSTOMER

The Dataset

What's in it

Everything is grounded in a synthetic Voice-of-Customer corpus: 12 fictional B2B SaaS accounts, each with a segment, ARR and health score, and 513 customer verbatims spanning reviews, survey responses, and win/loss and churn notes. It's entirely fictional, with no real customers, generated with Claude and embedded with Voyage, but shaped to feel real, including a few “at-risk” accounts whose feedback turns negative over time.

Explore the raw data