live system Live System Live Demo + Iteration

Steve — Voice Agent Sandbox

Public Runtime + Recap Pipeline

A live voice-agent sandbox for testing runtime UX, recap pipelines, follow-up logic, and public-facing conversational AI behavior.

Back to Projects

Project Brief

5 minutes

Demo window

Signed URL

Session boot

Recap pipeline

After the call

01 - Project Brief

Problem, Hypothesis, Outcome.

TYPE · STRUCTURED CASE NOTE

Summary

A public voice-agent project with signed access, recap lookup, transcript handling, rate limiting, and a private regression harness behind the scenes.

Problem

Most voice demos stop at the call itself and ignore the messy system work that starts before and after the conversation.

Hypothesis

A voice agent only becomes believable when the runtime, recap, privacy, and follow-up layers all work together under real use.

Outcome

Built a public sandbox that shows the runtime and the recap pipeline together, so the proof is not just the call but the whole system around it.

02 - Goals & Stack

What the build was trying to do.

TYPE · GOALS / TOOLS

Goals

Give people a real runtime they can actually try.
Show how recap and follow-up states behave after the call ends.
Keep the public surface constrained enough to be safe to expose.

Technologies Used

ElevenLabs Astro Cloudflare Signed session URLs Recap pipeline Rate limiting

03 - Breakdown & Notes

Implementation notes.

TYPE · DOCUMENTATION

Breakdown

Steve exists to prove that voice AI is a systems problem, not just a prompt problem. The public surface handles session launch, privacy framing, rate limiting, and the call itself. After that, the recap layer has to resolve correctly, present the right post-call state, and make human follow-up possible.

The public demo is only one layer. Behind it is a separate harness for scenario testing, which makes it easier to rehearse edge cases without turning the public page into a chaotic lab bench.

Build notes

Session launch uses signed access rather than exposing direct credentials.
Runtime and recap handling are intentionally separated.
Post-call states matter because they prove the call produced something usable.
The private simulation harness is what keeps the public surface honest.

Lessons Learned

The hardest part of voice AI is not the voice itself. It is everything around the voice: recap quality, privacy boundaries, handoff design, and whether the system is still coherent after the conversation ends.

04 - Analysis

Findings.

TYPE · FINDINGS

Live public demo that people can actually use.

Structured recap states rendered after real sessions.

Architecture story tied to runtime, recap, evaluation, and follow-up.

05 - Live Demo

Try the runtime.

TYPE · OPTIONAL LIVE SURFACE

This project includes a live public demo. The page still uses the same structured case-study format, but the runtime is available if you want to test the system directly.

Steve

Voice Systems Demo

Ready

Latency --ms

Window 5:00

State Standby

Real-World Demo

Steve has 5 minutes to understand the use case.

Start with the system or workflow you are exploring, what breaks today, and what outcome you care about. Steve will map the context, capture the signal, and show how the recap pipeline behaves under a real conversation.

You share

Project context, workflow details, and the friction in the current system.

Steve does

Maps the use case, surfaces the constraints, and points to the next step before the window closes.

[ Connect ]

Worth a conversation?

If you want to talk through Steve itself, conversational runtime design, or what it takes to turn a voice demo into a real system, reach out.

All Projects →

You are reaching

John Meyer

Security Engineer → AI

Open to roles
Contract + consulting
Architecture advisory