Sector
Energy retail, enterprise voice AI
Selected work · Case study
Conversational design at enterprise scale.
The prompt architecture, design tooling, and governance behind production voice agents on Amazon Connect and Nova Sonic 2.
The problem
A national energy retailer runs several voice assistants in production: identity verification, pay-as-you-go top-up, and a hardship-support credit journey. The brief was never to make a chatbot sound better. It was to build the system underneath assistants that verify identity and apply account credit, where a fluent wrong answer is worse than a stilted right one.
Scope of work
Prompt engineering, design tooling, and the governance that connects them to production.
- Prompt architecture & orchestration (Nova Sonic 2)
- Cross-team governance: security, data, product
- AI-assisted design tooling (Gemini)
- Call-transcript-driven validation
Canvas. A whole journey on one page. Node palette and reusable patterns on the left, global rules and variables on the right, live lint status underneath.
Multiple pages. Related journeys live in one shareable flow. Each page has its own canvas, and rules are inherited across all of them.
Review. Ask for a change in text or voice. The assistant returns a proposal, not an edit, and nothing changes until the designer chooses Apply changes.
Preview. Click through a path and read the transcript a caller would hear before anything reaches engineering.
Export. One flow, deterministic outputs: system prompt, tool configuration, sample dialogues, Lex V2 intents, and test cases.
Screenshots of the tool itself, using an anonymized demo flow. No client data.
Role
Prompt architecture, cross-team orchestration & governance
Stack
Gemini/Vertex AI · Amazon Connect · Nova Sonic 2
What it does
A design tool that sits between Figma and a build system. Conversations are assembled from reusable components (prompts, intents, conditions, actions) and used from client demos through to developer handoff. It is built around Amazon Connect and Nova Sonic 2, where hand-building a voice-to-voice flow is slow and easy to get wrong.
At enterprise level a voice-agent requirement can reach production as a full feature in about a sprint. Smaller deployments, such as a hotel or a single-location service, move faster still. I built it for my own project and it has had good feedback inside the company.
- Less time to build a voice agent
- Control over conversation and UX design
- Optimised, repeatable test cases
- Integration with Amazon services
My role
Not a UX pass on top of someone else’s system.
- Wrote the orchestration prompt architecture
- Aligned product, security, and data teams on what the assistant may do and see
- Built the tooling so designers can iterate without waiting for an engineering cycle
- Governed the result end to end, from what a prompt may read to what is recorded before anything ships
The model’s output is a reviewable diff inside a deterministic system, not the system itself.
How it fits together
Four parts of the system
01 — The tool
The assistant proposes edits rather than generating flows. Each proposal is limited to five allowed operations (no deletes) and is checked against the schema, the graph rules, and the linter before it is shown, then checked again before it applies. Nothing changes until the designer accepts, and a single undo reverses it.
02 — Shared skill design
Prompting rules are written down as versioned skills instead of living in one person’s head. Each rule belongs to exactly one layer of a seven-layer prompt architecture, and each traces back to a real call that went wrong. A fix lands once, and the next person, or agent, starts from what the team has already learned.
03 — Enterprise governance
Decisions move through four states (proposed, confirmed, implemented, deployed) in a shared ledger that names who confirmed each one. The tool’s export feeds the artifacts engineering imports, and an integration that is not built yet blocks publishing instead of passing silently.
04 — Research & validation
Every prompt version is reviewed against real call transcripts for behaviour, variance, latency, and guardrail gaps. A linter turns those manual findings into a check that fails on regression, so the same mistake does not ship twice.
Live demo
Try the tool
The anonymized showcase runs in the browser. No account and no client data. Explore the canvas, the reusable patterns, and the inspector rules, then open the preview and export.
Outcome
Faster first drafts without leaving a structured design model. Prompting knowledge that survives changes in the team. Decisions made once and recorded, so they are not re-argued by whoever opens the file next. Together they turn conversational design from a craft that lives in one person’s head into something a team can run.
Building voice agents on Amazon Connect? Let’s talk.
Book a call