Two workloads, two clocks
A spoken AI conversation needs immediate turn-taking; transcript analysis and summarization may run for seconds or minutes. Both have to coexist without blocking each other.
FIELD NOTES / CONVERSATIONAL AI SAAS
An AI that joins the call.
Not one that reads the notes.
BrieflyAI lets users configure an AI expert in plain English, meet with that agent through a live video call, and automatically receive a structured summary, speaker-attributed transcript, recording and contextual follow-up chat after the session.
CONVERSATIONAL AI CASE STUDY
BrieflyAI
Live AI voice participation.
Durable transcript processing.
BrieflyAI / Live AI meetingACTUAL PRODUCT INTERFACE
TWO CLOCKS, ONE PRODUCTLive turn-taking. Durable processing.
A spoken conversation needs immediate turn-taking; transcript analysis may run for minutes. Neither blocks the other.
Follow the lifecycle ↗ConfigureScheduleMeetTranscribeSummarizeAsk
DOMAIN
Meeting intelligence
WHAT WE BUILT
Real-time voice AI meeting platform
OUR ROLE
Product architecture to full-stack engineering
01 / THE REAL ENGINEERING PROBLEM
Answering a question is easy.
Running a meeting is not.
Building an AI that can answer questions is relatively straightforward. Building an AI participant that operates reliably across the complete lifecycle of a real meeting is a much broader product and engineering challenge.
Structured sessions need subject-matter experts, tutors, interviewers or trainers to repeatedly perform similar conversations. As demand grows, organizations either hire more experts or restrict how many sessions they can offer.
And when the conversation framework exists only in an employee's knowledge, different people deliver it differently — so the brief needs to be encoded into the system itself.
The browser is not the source of truth
Call start, call end, transcript availability and recording completion all arrive asynchronously from the media provider — and may be delivered more than once.
Knowledge dies in the recording
Explanations, decisions, commitments and questions get trapped inside recordings, or disappear entirely when no structured notes exist.
02 / INSIDE THE PRODUCT
Write the brief.
Meet the agent.
Six views of the delivered platform.
Each shows another stage of the meeting lifecycle.
Agents and meetings in one workspace.
The dashboard brings configured agents and scheduled sessions together, so the same place that defines a conversation also runs it.
THE MEETING LIFECYCLE
Configure→Schedule→Lobby→Converse→Process→Review
03 / REAL TIME AND DURABLE
The live call cannot wait.
The processing cannot be lost.
The two halves of the product have opposite requirements. Live conversation is latency-critical and cannot pause for analysis; transcription and summarization are slow and must survive failure.
So the slow work is handed to a durable pipeline. A dropped webhook never blocks the conversation record, and never duplicates it either.
Look under the hood ↘01
Hold the conversation live
Real-time voice AI participates in the call with immediate turn-taking, under the agent identity the brief defines.
Real-time voice
Turn-taking and speech happen at conversation latency, with the media infrastructure carrying the call.
Durable processing
Transcript, summary and recording are produced by background jobs that can retry without duplicating the record.
02
Hand off the slow work
Transcript processing and summarization run in a durable pipeline rather than in the request path of the call.
03
Reconcile the events
Meeting lifecycle events arrive asynchronously and can repeat, so processing is built to tolerate both.
04 / THE TECHNOLOGY & THE REASON
One product.
Two very different clocks.
The implementation keeps live conversation and post-meeting processing in separate execution paths, so the latency requirements of one never become the constraint of the other.
N
WEB APPLICATION
Next.js + React
Carries the workspace: agent configuration, meeting scheduling, the lobby, and the post-meeting review experience.
Why one application?
The place that defines a conversation is the place that runs and reviews it, so the lifecycle never leaves the product.
TYPED CONTRACTS
TypeScript + tRPC
End-to-end typed procedures between the client and server, so the shape of meeting and agent data is agreed in one place.
Why typed end to end?
Meeting state moves through many stages; a contract mismatch is far cheaper to catch at build time.
SV
MEDIA INFRASTRUCTURE
Stream Video + Chat
Runs the live video call and the chat surface, and emits the lifecycle events the platform reconciles against.
Why a media provider?
Call quality and delivery are their own engineering discipline; the product's value is the agent and the record, not the transport.
AI
LIVE VOICE
OpenAI Realtime
Gives the agent its spoken presence in the call, with the turn-taking latency a real conversation requires.
Why realtime specifically?
A participant that pauses to think is not a participant — live conversation and batch analysis are different problems.
IN
DURABLE PROCESSING
Inngest
Runs transcript processing and summarization as durable background work, outside the live conversation path.
Why durable jobs?
A dropped webhook never blocks the conversation record — and retries never duplicate it.
RECORDS
Neon PostgreSQL + Drizzle
Holds agents, meetings, transcripts and summaries as the durable record of every session.
Why a relational record?
A meeting is a set of related facts — agent, participants, transcript, summary — and they are queried together.
The architecture and product narrative reproduce the supplied project documentation. Interface values are product screen data.
Explore the documented platform stack
- Web application
- Next.js · React · TypeScript · tRPC
- Real-time media
- Stream Video · Stream Chat · meeting lifecycle events
- Live voice AI
- OpenAI Realtime
- Durable processing
- Inngest background jobs · retry-safe transcript and summary pipeline
- Data
- Neon PostgreSQL · Drizzle ORM
Technology names identify the documented implementation, not partnerships or endorsements.
A CONVERSATION ABOUT YOUR MEETING WORKFLOW
Which conversation
keeps getting repeated?
Bring one structured session your team runs over and over — tutoring, interviewing, intake, onboarding. In 20 minutes, we will map what an AI participant could hold, and what still needs a person.
20 minutes · Your session format and constraints