Proppify
Case study · Emergence Mirror Labs

Two governed AI companions, live to the public

Emergence Mirror Labs needed AI companions that could hold a real conversation without inventing guidance, and handle personal disclosure responsibly. We engineered and deployed both on ANT.

The problem

What needed solving

Emergence Mirror Labs works in transformation and reflective practice — a domain where a confidently wrong answer does real harm. A general-purpose chatbot was not an option: it would improvise guidance that the organisation had never approved, in a register that didn't match how they work.

The second constraint was the nature of the conversation itself. People disclose personal material to a companion like this, which makes session handling, consent and retention a design problem from the first line of code rather than a compliance checkbox at the end.

The approach

How we built it

We built both companions — ANI and Transformation Talks — on ANT, our agentic platform, rather than starting from scratch. That gave us grounding, retrieval, handoff and session handling as platform behaviour, so the work went into the parts that are specific to EML.

Each companion answers from a curated knowledge base of approved material. Where a question falls outside what has been approved, the agent says so and offers a handoff instead of filling the gap with plausible text. Getting that refusal behaviour right took more iteration than the answering behaviour did.

Human-in-the-loop escalation is part of the conversation flow, not a fallback bolted on afterwards. Consent and session-data handling were designed against Malaysia's PDPA requirements from the outset.

The stack

What it runs on

ANT agent platform
Retrieval over curated knowledge bases
Human-in-the-loop handoff
PDPA-aligned session and consent handling
Web delivery
Delivered

What shipped

Two companions, ANI and Transformation Talks, live to the public at emergencemirrorlabs.org.

A knowledge base pipeline the EML team can keep current, so answers don't quietly go stale after launch.

Grounded answering with explicit refusal and handoff paths where a question needs a person.

Consent and session-data handling built into the conversation flow.

Ongoing operation of both companions post-launch.

Have a similar problem? Let's talk.