Rizmataz — a voice-first founder network
Record a pitch once. Let your voice live on as replayable threads, live AI conversations, and memory-aware follow-ups.
Client — Rizmataz · Voice-first social network for founders
Rizmataz is a live product built by Agency7 — a social network where founders record pitches in their own voice, chat live with AI-powered founder agents via ElevenLabs, and carry thread memory forward between conversations. The entire stack is edge-first: Next.js 15 on Cloudflare via OpenNext, D1 for structured data, R2 for audio, and an OpenAI pipeline for transcription, summarization, and thread memory.
What is Rizmataz?
Rizmataz is a voice-first social network for founders. Instead of another text feed, it is built around the human voice — record a pitch once, save it to a reusable founder profile, and watch it evolve through live AI-powered conversations, threaded voice replies, and memory-aware follow-up chats.
Every founder has a persistent profile — name, bio, goals, up to three public links the AI can enrich from — and every pitch becomes a living object that others can reply to with their own voice notes, nested and replayable. The ElevenLabs-powered AI agent behind each founder chats back using a cloned voice, carries conversation memory forward across sessions, and produces new dated thread cards that anyone signed in can continue.
Where most social platforms optimize for scroll, Rizmataz optimizes for replay. A pitch recorded today is still talking tomorrow.
The challenge
Voice-first products are hard. You have to record cleanly on any device, transcribe accurately, score and summarize content that was never meant to be written, handle live bidirectional voice with under half a second of latency, clone voices without leaking identity, and persist conversation memory long enough that a returning user can pick up where they left off — even days later.
On top of that, Rizmataz needed to be globally fast, cheap to run, and resilient. A typical pitch is a 60–90 second audio clip that has to be transcribed, analyzed, stored, and instantly replayable from a dated timeline. Every live chat invokes a dynamic ElevenLabs agent with a signed URL, founder context, and a rolling memory window.
Our job was to design a stack that could handle all of that at the edge — not in a traditional region-bound cloud — while keeping the developer experience fast enough to iterate in public.
What we built
Founder profile & pitch recording
A dashboard flow where founders save a reusable profile (name, bio, feedback goal, up to three public links) and record a pitch directly in the browser. The pitch is uploaded to R2, transcribed by OpenAI, scored, and summarized automatically.
Live AI founder chat (ElevenLabs)
A one-tap live conversation with an ElevenLabs dynamic conversational agent. Signed URLs, per-session context injection, and real-time bidirectional voice so founders can continue pitching to their own AI version — or to someone else's.
Voice cloning
A voice-clone endpoint wired to ElevenLabs so each founder's AI agent speaks in their own voice. Clone once; every future agent response reuses the cloned voice ID for continuity.
Threaded voice-note replies
Nested voice replies on every pitch and every conversation card. Anyone signed in can record, reply, and replay — building a living thread around each pitch.
Dated conversation timeline
Every live chat creates a dated conversation card that anyone can return to, replay, and continue. The timeline is the product's memory.
Thread memory
OpenAI-generated summaries of each thread get folded into the next conversation's context, so the AI remembers what was said in prior sessions and continues naturally.
Link enrichment (optional)
Firecrawl-powered scraping of a founder's public links gives the AI richer context — website copy, recent posts, press — so the cloned agent can speak intelligently about the founder's actual work.
Public feeds
Public pitch feed, public founder feedback feed, and a gallery surface discoverable content for signed-out visitors — the social layer on top of the voice substrate.
How we built it
Frontend
- —Next.js 15 — App Router, TypeScript, React 19
- —Tailwind CSS + shadcn/ui — Radix primitives for accessible components
- —React Hook Form + Zod — type-safe form + validation
- —TanStack Query — client-side data sync
Edge infrastructure
- —Cloudflare Workers — two workers — app worker (OpenNext-served Next.js) + data worker (custom routes)
- —OpenNext for Cloudflare — Next.js → Workers runtime adapter
- —Cloudflare D1 — pitches, founder sessions, thread metadata
- —Cloudflare R2 — pitch audio, voice-note media, worker cache
- —Wrangler — deployment, bindings, schema migrations
AI & voice
- —OpenAI — transcription, pitch scoring, thread summaries, conversation memory, artifact generation
- —ElevenLabs Conversational AI — dynamic agents, signed URLs, voice replies
- —ElevenLabs Voice Cloning — per-founder voice IDs reused across agents
- —Firecrawl — optional scraping to enrich founder context from public links
Auth & monitoring
- —Clerk — authentication, session management, route protection
- —LogRocket — session replay, error tracking
- —Plausible — privacy-friendly analytics
The user flow
Everything on Rizmataz hangs off one clean spine: founder profile → pitch → live chat → dated thread → nested replies → memory. Each step is a single Next.js route backed by a Cloudflare Worker data endpoint.
- 01
Save founder profile
A one-time Clerk-protected flow captures name, bio, feedback goal, up to three public links, and a default pitch title. Profile data lives in D1.
- 02
Record a pitch
The browser records directly to R2 through a signed upload. A route handler calls OpenAI for transcription, scoring, and a summary, then stores everything against the founder's profile in D1.
- 03
Start a live founder chat
One tap spins up an ElevenLabs dynamic agent with a signed URL and founder context (profile + pitch + enriched links). Bidirectional voice opens in-browser with sub-second latency.
- 04
Create a dated conversation card
Every chat produces a timestamped card on the founder's timeline. The full audio and transcript are persisted to R2/D1 so the thread is replayable forever.
- 05
Nested voice-note replies
Anyone signed in can record a voice reply on any card — recursively. Each reply is its own R2 object with its own transcript; the timeline renders them as a true tree.
- 06
Thread memory folded forward
An OpenAI summarization step compresses each thread into a compact memory chunk. The next ElevenLabs chat receives that memory in its dynamic context — so the agent remembers what was said before, even weeks later.
Outcomes & learnings
- Shipped a production voice-AI product on a Cloudflare-first stack — globally distributed, under a second from cold load on most devices.
- ElevenLabs dynamic agents with per-session signed URLs scaled cleanly across users without per-founder backend infrastructure.
- OpenAI thread summarization reduced memory payloads by roughly 90 percent vs raw transcripts while preserving enough context for coherent follow-ups.
- R2 + D1 handled pitch audio and threaded metadata at edge latency with no region-bound database to scale.
- Two-worker architecture (app + data) proved out as a repeatable pattern — we reuse it on other Agency7 voice-AI projects now.
- Validated that voice memory, not voice recording, is the hard part. Getting summaries, retrieval, and continuity right is what makes a voice product feel like it remembers you.
Want a voice agent for your business?
Rizmataz pushes voice AI to its hardest case — live, memory-aware, publicly threaded. But the same building blocks (ElevenLabs conversational agents, OpenAI transcription and memory, edge infrastructure) are what we deploy for Edmonton service businesses that want an AI voice agent answering their phone 24/7.
AI voice agents in EdmontonWant something like this?
Free 15-minute strategy call. We'll walk through what a similar system could look like for your business.
Book a free AI voice discovery call