← Back to Bulkhead Forge

Bulkhead Forge — product / Actions

Meetings go in. Tracked action items come out.

Actions listens to the conversations your team already has — Google Meet today, Slack and email next — transcribes them, and turns every decision into an assigned, deduplicated action item. A tracker that fills itself.

Request early access See the pipeline

Working MVP · Runs on our own Kubernetes cluster

Illustrative pipeline output — the stages are real, the file and counts are placeholders, not a measured client run.

The problem

Meetings end. Follow-ups evaporate.

Decisions get made in calls, then live in one attendee's notes — if anywhere. Trackers only know what someone bothered to type in, so they drift from reality within a sprint. Actions closes the gap at the source: the conversation itself.

How the same three artefacts fare today versus with Actions in the loop.
Artefact Today With Actions
Notes One attendee's doc, unread after Friday Structured analysis saved with the meeting
Follow-ups Re-typed into a tracker by hand — or not at all Extracted, assigned and deduplicated automatically
Decisions Folklore within two weeks Logged with the meeting, the owner and the date

How it works

From spoken word to tracked item

  • Phase 01

    Capture

    New Google Meet recordings are detected automatically. Any video or audio in Drive can be submitted by hand. Typed items work too — automation shouldn't forbid keyboards.

    Meet · Drive · manual

  • Phase 02

    Transcribe

    Audio is extracted with ffmpeg, staged to Cloud Storage and transcribed by Speechmatics or Speech-to-Text V2 (Chirp 3) — a console setting picks which. The transcript is saved with the meeting.

    ffmpeg · Cloud Storage · Speechmatics / STT V2

  • Phase 03

    Analyse

    A model reads the transcript under an enforced JSON schema: summary, key topics, decisions, action items, open questions, risks, follow-ups. Constrained output — not free-form parsing.

    Structured output

  • Phase 04

    Track

    Items land in the console with owners, labels, comments and subtasks. A confidence-gated second pass catches duplicates — a re-raised topic gets a "discussed again" comment, not a clone.

    Dedup · high / med / low

What you get

Tracker ergonomics, minus the typing

  • Auto-extracted action items

    Owner, status, due date — pulled from speech, then editable like any task you'd have typed.

  • Duplicate guard

    Dedup verdicts carry high / medium / low confidence; you set the weakest verdict that's allowed to act.

  • Decision log

    Every decision is kept with its source meeting — the "why" survives the quarter.

  • People, not strings

    Speakers resolve to real people; assignees are accounts with history, not names in a cell.

  • Context that survives

    Labels, comments, subtasks and relations between items — everything a tracker owes you.

  • Full activity trail

    Who changed what, when — every item carries its history, human edits and agent edits alike.

Sources

A source is an adapter

The pipeline doesn't care where the words came from. Each channel is a thin adapter in front of the same transcribe → analyse → track path — so adding one doesn't touch the core.

Live today

  • Google Meet — recordings detected automatically
  • Google Drive — drop any video or audio file
  • Manual — create and edit items by hand, anytime

On the roadmap

  • Slack — threads and huddles into the same pipeline
  • Zoom — recordings from outside the Google world
  • Email — decisions made in writing count too

Your team talks somewhere else? Tell us — adapters are the cheap part.

Infrastructure

Runs on our own cluster

On our cluster

  • Services and pipeline jobs on k3s, on a server we operate
  • Deployed from Git by Argo CD — a merge is the deploy
  • PostgreSQL (CloudNativePG) and Vault for secrets
  • Langfuse — meeting analysis: prompts versioned, model calls traced

Services it uses

  • Google OAuth · Drive & Meet APIs
  • Speechmatics or Speech-to-Text V2 (Chirp 3), chosen in the console
  • Models through OpenRouter, one per role, chosen in the console
  • Secret Manager · Artifact Registry · keyless WIF deploys

Roadmap

Where this goes

Now

  • Meet auto-ingest & Drive uploads
  • Analysis, dedup, decision log
  • Console tracking: people, labels, comments

Next

  • Slack & email adapters
  • Slack notifications on new items

Later

  • Two-way sync with Jira & Linear
  • Cross-meeting insights per person & topic
  • Public API for custom sources

Bulkhead Forge

Three products, one platform

Actions

Meetings become assigned, deduplicated action items. You are here.

 Discovery

Standing questions answered on a schedule, with evidence attached.

 Agent Delivery

A control plane for agent-driven development — gates, drift guard, budgets.

Early access

A tracker that fills itself. Try it before everyone else.

Working MVP, onboarding a first cohort of teams on Google Workspace. No forms — one email is the whole first step.

Request early access Book a free architecture review

contact@bulkheadlogic.com