Work Podcast platform

Helsekoden

Services
Podcast platform, Search, Community, MCP
Year
2026

Four hundred episodes, and every word of them findable.

Helsekoden is AFPT's podcast about training and health. We built its home: every episode written out, searchable to the second, and open for the listeners to answer back.

Where it started

A feed and a Facebook group

The podcast lived in a feed: you could listen, and that was all. Questions from listeners came in through a Facebook group and sank down the page. To find an episode again, you had to remember its title.

The show needed a home where what had been said could be found, and where the listeners had somewhere to answer.

What we built

Everything said, findable

  1. Find what was said

    Before: an episode could only be found by its title.

    Now: the search reads everything said in every episode, by meaning, and opens the episode at the second it was said.

  2. Answer at the right moment

    Before: a reaction to something said had no place to go.

    Now: a comment is pinned to a second of the audio, and sits as a marker on the sound itself.

  3. The hosts, by what they said

    The need: a show with years of episodes knows very little about itself.

    What we did: who talks the most, and which words give each host away, come straight out of the transcripts.

Listener questions

Questions that stay answered

Before: listener questions were scattered across Facebook posts, and an answer given on air was gone once the episode had played.

Now: every question is in an archive, linked to the episode and the second where it was answered. The old episodes were gone through too.

Behind it

One decision per episode

The need: four hundred episodes to write out, sort and make searchable, and a new one every week.

What we did: a new episode is fetched from the feed and sent to a transcription service, which writes it out and tells the voices apart. A person says who is who. After that the text is cleaned up, cut into passages and turned into vectors, so a search goes by what was meant and not by which words were used.

Below: one episode on its way from the podcast feed to a search, and what a search does.

A new episode

The podcast feed, to the team

New episode: Sleep, evenings and small habits.

Who is speaking?

Voice A

Voice B

Ready

To the team

The episode is ready. Listeners can now search everything that was said in it.

Helsekoden's platform

  1. Sent to be written out

    Sends
    A link to the sound, not the file
    Gets back
    Every word, with its time and its voice
  2. The ads are told from the talk

    Marks
    The sentences that are ads
    So that
    They stay out of the search
  3. A person says who is who

    Names
    Each voice as a host or a guest
    Then
    Nothing more is asked of anyone
  4. Cleaned up by fixed rules

    Fixes
    Names and terms it mishears
    How
    By rule, never by guessing
  5. Cut into passages, and embedded

    A passage
    One speaker's turn, with who said it and when
    Kept as
    Vectors in the platform's own database
  1. AssemblyAI

    Turns speech into text, and tells the voices apart

    • Is written out
  2. A language model

    Reads the opening minutes and marks what is an ad

    • Ads are marked
  3. The clean-up rules

    What the transcription keeps getting wrong, and what it should say

    • Is cleaned up
  4. OpenAI embeddings

    Turns a passage into a vector: numbers that stand for what it means

    • Is made searchable

Naming the voices was the one thing a person did.

A listener searches

Search on helsekoden.no

What did they say about magnesium in the evening?

The answer

The guest, in “Sleep, evenings and small habits”: “I hear about magnesium all the time too. For me the routine matters more.”

Play from 18:40

Helsekoden's platform

  1. The question becomes a vector too

    Compared by
    What it means, not which words it uses
    So
    A question in other words still finds it
  2. The nearest passages are found

    Looks through
    Every published episode
    Keeps
    The passages that come closest
  3. An answer, from what was said

    Uses
    Only the passages that were found
    Shows
    Who said it, and the minute
  1. OpenAI embeddings

    Turns a passage into a vector: numbers that stand for what it means

    • Is read for meaning
  2. The vector index

    Every passage of every episode, in the platform's own database

    • Finds the passages
  3. A language model

    Writes a short answer from the passages it is given

    • Answers from them

Every result opens the episode at the minute it was said.

MCP

What happens when the team asks their assistant

To bring in what is new

Helsekoden's team, in their own AI assistant

Check the feed and bring in what is new.

Back in the assistant

The new episodes are in. Writing them out is a step of its own, and I will ask before I start it.

The staff server on helsekoden.no

  1. Signed in as themselves

    Who
    One of Helsekoden's team
    May
    What they may do in the admin, nothing more
  2. Says what it will do, and waits

    Would
    Bring in what is missing from the feed
    Needs
    A yes
  3. The new episodes are brought in

    Adds
    The episodes that were missing
    Logged
    In their name
  1. Roles and permissions

    What each person may do in the admin

    • Is recognised
  2. The podcast feed

    Where a new episode first appears

    • Shows it first
  3. The log

    Every change, and who made it

    • Brings them in

Nothing that costs money starts without a yes.

Who needs a speaker check

Helsekoden's team, in their own AI assistant

Which episodes still need a speaker check?

Back in the assistant

A list of the episodes that wait for someone to say who is who. That is the one step a person does.

The staff server on helsekoden.no

  1. Signed in as themselves

    Who
    One of Helsekoden's team
    May
    What they may do in the admin, nothing more
  2. Looked through the episodes

    Finds
    Written out, with a voice nobody has named
    Changes
    Nothing
  1. Roles and permissions

    What each person may do in the admin

    • Is recognised
  2. The episodes

    Every episode, and how far it has come

    • Looks through the episodes

Only a read. Nothing was changed.

To clear out the spam

Helsekoden's team, in their own AI assistant

Show this week's comments and take out the spam.

Back in the assistant

Three comments taken out, one at a time, each after you said yes.

The staff server on helsekoden.no

  1. Signed in as themselves

    Who
    One of Helsekoden's team
    May
    What they may do in the admin, nothing more
  2. This week's comments, read

    Shows
    Everything written since Monday
    Changes
    Nothing yet
  3. Each one shown before it is taken out

    Shows
    Who wrote it, and what it says
    Needs
    A yes for each one
  1. Roles and permissions

    What each person may do in the admin

    • Is recognised
  2. The comments

    What listeners have written under the episodes

    • Reads the comments
  3. The log

    Every change, and who made it

    • Shows each one first

Logged in their name, like everything else they change.

What changed

What we solved for the podcast

The archive is something the show can use now, not something it has. And the listeners have a place of the show's own to talk in.

  • Every episode can be found by what was said
  • Listeners answer at the exact moment
  • Questions keep their answers
  • The conversation is on the show's own site
  • A new episode takes one decision
  • Staff run it by asking

Let's talk

Tell us what you want to build, and we will work it out together.