Skip to content
Relay.

01 · RELAY / 2026

RELAY.AGENT
listening for new tasks

Meetings
run themselves.

An AI meeting automation layer that plugs into your PM tool. Built for Asana. Ready for any tool.

THE SYSTEM · ONE GLANCE

From a meeting to a filed ticket.

Six stops. Zero human steps in between.

01MEETING COORDINATION

BD sets the time.

Three fields in Asana — time, platform, client email. That's the only manual step in the whole system.

02MEETING COORDINATION

A meeting link spawns.

Relay detects the update and creates a Zoom or Google Meet link automatically — no one has to hit a button.

03THE CALL

Human-led, bot in back seat.

Your team runs the call. A recording bot joins silently and captures everything for later.

04AI PROCESSING

Agent 01 summarizes.

The Summarizer turns the transcript into a structured note — overview, decisions, action items, risks.

05AI PROCESSING

Agent 02 estimates.

The Estimator reads the summary and returns hours, budget, and timeline — ready to paste into a proposal.

06BUSINESS OUTPUT

A ticket, filed.

Everything lands back in Asana: meeting link, transcript, summary, cost estimate, and a child task for every action item.

02 · THE PROBLEM

Before Relay, a meeting cost you seven steps.

BEFORE

  • 01.BD confirms the meeting time.
  • 02.Someone manually creates a Zoom link.
  • 03.Someone sends the calendar invite.
  • 04.Someone remembers to invite Fireflies.
  • 05.Someone goes hunting for the transcript.
  • 06.Someone writes the summary by hand.
  • 07.Someone updates the Asana task.

AFTER

BD sets a meeting time.

— that's it.

03 · THE PIPELINE

A task becomes a transcript.

Use the arrows, tap the dots, or swipe to step through.

01 / 06

A task is born.

BD fills three Asana fields: confirmed time, platform, client email. Nothing else.

01 / 06

04 · UNDER THE HOOD

Under the hood.

One central system coordinates everything. A database tracks every step — what's done, what's pending, what failed — so nothing happens twice and nothing gets lost. If a service hiccups mid-flow, the system resumes exactly where it left off.

YOUR TASK BOARDAsana · WebhookCENTRAL SYSTEMNestJSlistens · schedules · retriesDATABASEPostgreSQLtracks every stepZOOM / MEETMEETING RECORDERFirefliesLOCAL AI MODELOllama · Llama 3.2BACK TO YOUR BOARDAsanacreate meetinginvite bot · get transcriptsummarize on your serverwrite results back
NestJSTypeORMPostgreSQLZoom S2S OAuthGoogle CalendarFireflies GraphQLOllamaLlama 3.2

05 · HARD PROBLEMS SOLVED

The hard parts.

Automation looks easy until webhooks fire six times, meetings overlap, and the LLM times out. Here's what we handled.

01IDEMPOTENCY

Webhooks fire many times. We process once.

Asana sends the same task-updated webhook repeatedly. We read boolean flags in Postgres — meeting_created, transcript_fetched, summary_generated — as the source of truth. Never a duplicate call.

02CONCURRENCY

Unlimited parallel meetings. No collisions.

A unique constraint on asana_task_id in the meetings table. The scheduler processes each row independently. Two meetings at 2 PM? Two rows, two transcripts, two summaries, zero cross-talk.

03FAULT TOLERANCE

Zoom down? Fireflies lagging? We resume.

Exponential backoff on every external call. The scheduler is idempotent and re-queries Postgres on each tick. Server restarts mid-pipeline? Next cycle picks up where we left off.

04LOCAL-FIRST AI

Transcripts never leave your server.

Ollama runs Llama 3.2 on the same box as the NestJS service. No OpenAI, no Anthropic, no cloud inference. Privacy and cost benefits — and it runs on commodity hardware.

06 · INSIDE AN ASANA TASK

What a client actually sees.

10:00 AMBD FILLS THE INPUT FIELDS
2:20 PMRELAY FILLS THE REST

No human touched this task between 10:00 AM and 2:20 PM.

07 · BEYOND THE SUMMARY

More than notes.

The transcript is just the start. Once the discovery call ends, Relay can turn the conversation into real work — not just a memory of it.

01

Project estimates.

Hours, budget, and a realistic timeline — drawn from the call alone. Drop it straight into your proposal.

02

Action items, pre-assigned.

Every action item becomes a child Asana card, already assigned to the person who volunteered on the call.

03

Follow-up drafts.

A ready-to-send email summarizing what the client heard and what happens next — so nothing stalls after the call ends.

04

Risks, flagged early.

The AI spots scope creep, contractual edge cases, and timeline conflicts the team might have missed in the moment.

+ plug in any prompt. the pipeline already has the transcript.

08 · What it speaks to

Asana·Zoom·Google Meet·Fireflies·Ollama·NestJS·PostgreSQL·Asana·Zoom·Google Meet·Fireflies·Ollama·NestJS·PostgreSQL·

09 · CONTACT

Want something like this built?

Relay was built by Silicon Pixels, a small studio that ships production-grade backend systems, AI integrations, and automation pipelines. We like problems that look simple from the outside and turn out to hide seven edge cases underneath: webhooks that fire six times, LLMs that time out, meetings that overlap, transcripts that arrive late. If you want something like Relay built for your workflow, let's talk.