Employment · Software Architect at TED · Platform architecture
TED.
Four and a half years on a product millions of people use.
The talk player, saved playlists, search, and the advertising that keeps the talks free.
01 / The problem
What needed
to change.
ted.com has to do several things at once. Be quick for someone who just wants to watch a talk, findable for someone searching, worth returning to, and profitable enough to keep the talks free. Work on any one of those can quietly damage the others, and most of the damage is invisible until someone goes looking.
02 / What changed
The work,
and what it changed.
Advertising pays for a lot of what TED gives away, and revenue fell sharply with no obvious cause. The failures left no trace anyone could see.
Instrument the ad stack before changing it, so every slot reported what it actually did, then fix the startup race the data exposed and retire slots that never filled.
I led the team that traced the loss and recovered it. The stack now reports its own health instead of failing quietly.
The player is the product. Advertising runs inside it, and every stumble there is a stumble in the thing people came for.
Load the talk's own video first, and make the handoffs to advertising explicit so each side knows when the other starts and stops.
The talk begins on its own terms, and the ad breaks stopped stepping on it.
People could watch talks but had no way to keep the ones that mattered, or hand a set to someone else.
Build saved talks and personal playlists end to end, from the save dialog through to sharing, and release it behind a flag so it could be turned on gradually.
Visitors collect and share their own playlists, and a shared link works for someone opening it cold, with no account and no history.
Search is how most people find a talk, and the technology behind it was years out of date.
Rebuild search as its own service, and prove the new one returned the same results as the old one before anything switched over.
People kept finding talks throughout the change, with no moment where a search came back empty.
Also in those years
- Rebuilt the talk page experience behind a feature flag, so it could be turned on gradually rather than in one jump.
- Cut layout shift, cached the page banners, and built the dashboards the team uses to watch loading speed.
- Rebuilt the speaker and people directories on one filter pattern, with a phone-friendly panel, names that wrap properly, and correct indexing on filtered views.
03 / The result
What the work
made possible.
Four and a half years of work on a product people use every day: saved playlists they did not have before, a steadier player, a search that survived its own replacement, and an ad stack that reports its own health. This is team work done in my role at TED, alongside other engineers, design, and product.