People Hub

A multi-scene AI avatar video and companion landing page that turns Airtable's internal employee directory into an engaging, shareable product walkthrough.

5 days

Timeline

1:33

Video runtime

7

Scenes

Solo

Role — builder

A powerful tool that needed a proper introduction

People Hub is Airtable's custom-built internal employee directory — a full-featured platform built in Airtable itself, covering employee profiles, org charts, workforce analytics, skills tracking, and culture features like Cheers and Fun Stats. Getting employees to discover and use it required more than a wiki page.

Three interconnected deliverables

The project produced a named AI agent, a production-quality video, and a companion website — all built end-to-end in under a week.

Scene breakdown

Each of the 7 scenes pairs a real People Hub screenshot as the background with the avatar narrating that section of the script.

Real screenshots from the live platform

These are the actual People Hub screens captured during production — each one used as a scene background in the video. Personal and third-party details are redacted.

Four engineering problems, four clean solutions

The production wasn't straightforward. Each challenge led to a durable solution that now lives as a reusable skill or workflow pattern.

1. 01

   Community HeyGen skill failed security audit

   The off-the-shelf HeyGen skill from the community registry couldn't be used — it didn't meet the security requirements for a credential-holding skill in this environment.

   Fix:

   Built a custom heygen-custom-avatar skill from scratch using only Python stdlib (no third-party dependencies). It submits video generation requests to HeyGen's v2 API, polls for completion, and returns the MP4 URL. Now a reusable team asset.

2. 02

   SSO wall blocked browser screenshot capture

   The People Hub interface requires Airtable SSO login. The agent's browser automation couldn't authenticate through corporate SSO — every attempt hit the login wall.

   Fix:

   Pivoted to a user-provided screenshot workflow. I captured 13 real screens from my own authenticated session and dropped them into the thread. The agent then incorporated them as scene assets — producing more authentic footage than a browser-automated capture would have.

3. 03

   HeyGen rejected wrapper URLs — needed raw bytes

   The first multi-scene attempt passed stored-file URLs as background images. HeyGen's API received them but returned parse errors — because those URLs serve a wrapper page, not raw image bytes.

   Fix:

   Switched to GenerateTempExternalDownloadUrl, which returns S3 presigned URLs serving raw image/jpeg bytes directly. HeyGen parsed these cleanly. This pattern is now documented as the correct approach for any integration that fetches image assets by URL.

4. 04

   First render was 89 MB — over the 50 MB platform limit

   The initial script was a faithful 13-section walkthrough (~10 minutes spoken). The first video came back at 89 MB — well over the 50 MB upload ceiling.

   Fix:

   Condensed the spoken script to a ~1:33 target. The fuller narration moved into the companion webpage's editorial copy instead, so no content was lost — it just lives in the right medium. The video became tighter and more watchable as a result.

Built with

Every component was orchestrated through a single Hyperagent thread — no video editing software, no external production tools.

What this unlocked

The project delivered a reusable production pipeline — not just a one-off video.