Methodology
Filter before you spend. Measure before you claim.
Most short-form video carries no usable speech. The pipeline classifies that cheaply and early, so expensive work only ever touches material that will become a record.
Unstructured noise
Millions of videos a day, most of them carrying nothing usable.
Convergence / filter
Cheap classification first, so cost only lands on material that will become a record.
Structured signal
A flat, versioned table with a stable schema and the sample size on every row.
- 01
Video sources
Uniform sampling of the short-form video space, not keyword search. The frame is drawn before anything is labelled, so the sample carries a known denominator.
- 02
Transcription
Speech is separated from catalogue audio and reused sounds, then transcribed. Voice-activity detection decides what is worth transcribing before any compute is spent.
- 03
Entity & topic
Transcripts are resolved against a maintained taxonomy and a canonical entity list. Platform-native labels are used to include, never to exclude.
- 04
Stance & sentiment
Sentiment is scored −1 to +1 and stance is scored toward the entity, not the clip. The two are kept separate because a positive delivery can carry a negative stance.
- 05
Region & time
Records are attributed to a region and a time bucket, then aggregated. Aggregates carry the sample size and a 95% confidence interval.
- 06
Structured dataset
The output is a flat, versioned table with a stable schema — the same records behind the API, the Parquet drop and the CSV backfill.
Quality assurance
A number is only useful if you know what it is a number of.
Known denominator
Sampling is uniform over the identifier space, so a share is a share of a measurable population rather than of whatever a search returned.
Positive controls
Live records of known state are interleaved with every collection run, which separates 'absent' from 'blocked' instead of silently under-counting.
Sample size on every row
Every aggregate ships with n and a 95% interval. Rows thinner than the stated floor are published as null rather than smoothed.
Human validation
A stratified slice of each release is hand-checked against the transcript, and the agreement rate is published with the release notes.
Request a sample dataset.
Tell us the market and the question. We return a real slice — schema, sample sizes and confidence intervals included — not a brochure.