SKAPIAN LAB◆ CASE STUDY · 10 — MEDIA METADATA EXTRACTION
Structured tags. Live feed.
Categorical metadata over CNBC, Fox News, and more — not summaries. Filterable, sortable, aggregatable. Surfaced on one editorial dashboard.
Multi
Channels indexed
Live
Trending detection
Azure
End-to-end pipeline
10 wks
Pilot to production
— 01 / LIVE INGEST
Four channels. One incoming stream.
Articles and broadcast clips ingested in real time. Each item tagged with channel, timestamp, and routed into the extraction queue.
CNBC
Business · Finance
412 today
FOX NEWS
Politics · Current
284 today
ESPN
Sports · Entertainment
196 today
NPR
General News
138 today
INCOMING · LIVE● 1,240 INGESTED TODAY
14:32:08CNBCFed holds rates steady at 5.25%, signals two cuts in 2026● QUEUED
14:31:54FOX NEWSSenate hearing on AI safety convenes Tuesday — Altman, Hassabis to testify● QUEUED
14:31:32ESPNCity overturn United in Manchester derby — Haaland hat-trick● QUEUED
14:30:48CNBCApple raises full-year forecast — services hits record● QUEUED
14:30:12NPRCyclone Vora makes landfall on Andhra coast — evacuations…● QUEUED
— 02 / THE EXTRACTION ENGINE
One pass. Six structured fields.
Categorical extraction — not summarisation. Topic, sentiment, content-safety, trend, named entities, and category — all in one model pass, with the source document still hanging off the side.
— 03 / EDITORIAL DASHBOARD
Filter, sort, aggregate. Same as a spreadsheet.
Output lands in Postgres as structured rows. Editorial filters by topic, sorts by trend, aggregates safety flags — without writing SQL.
EDITORIAL · STRUCTURED OUTPUT
TOPIC ↓SENTIMENTSAFETYTREND ↑
HEADLINETOPICSENT.SAFETYTREND
Fed holds rates steady, signals two cutsMarketsNeutral✓↑ High
Senate hearing on AI safety TuesdayPoliticsMixed⚠↑ High
Manchester derby — Haaland hat-trickSportsPositive✓↑ Spike
Apple raises full-year forecastMarketsPositive✓→ Stable
Cyclone Vora makes landfall on Andhra coastWeatherNeutral✓↑ Local
Layoffs at major retailer hit 4,200MarketsNegative⚠↓ Falling
AGGREGATES· Content-Safe 94%· Trending 38· Negative-sentiment cluster +12
— 04 / END-TO-END PIPELINE
Scrape to schema. Continuously.
Scrape
Multi-channel feed
Normalise
Strip · structure
Extract
Azure OpenAI · GPT-4o
Validate
Schema · guardrails
Store
Postgres · indexed
Serve
Editorial dashboard
— 05 / IN PRODUCTION
Summaries out. Structure in.
1,240+
Articles / day extracted
6
Categorical fields · one pass
94%
Content-safe rate
10 wks
Pilot to production
"Editorial wanted to know what was trending, what was safe, and what was top-performing — in structured fields, not summaries they couldn't filter."
— The constraint that made this fit
— 06 / BUILT WITH
The toolkit.
Azure OpenAI
Azure Functions
Python
Postgres
— NEXT CASE STUDY