I build the systems that help entertainment move at the speed of culture.

When a client breaks through culturally, a viral post, a red-carpet look, a news beat, the department usually knows within a day. The value does not leak at detection. It leaks in the 48 to 72 hours after, when the team scrambles on rates and targets while the leverage window closes. By the time outreach is ready, the moment has cooled.

The real bottleneck was readiness, not detection. That reframe drove every decision that followed.

Collection runs outside the AI. A vendor-side scheduler scrapes each watchlisted client's public social activity daily and lands it in the data store. Press mentions arrive through free alerts. This layer deliberately does not depend on anyone's laptop being open or any single person's attention.

Judgment runs inside a daily AI pass. One scheduled analysis task fires each morning. It measures each client's engagement against that client's own trailing 30-day baseline rather than a fixed follower threshold, so a genuine spike registers at any audience size. It flags anything crossing the moment threshold, classifies the moment type, and matches it against an approved brand directory and a forward events calendar. It never invents brand names. It assembles a moment brief, numbers, moment type, best-fit brand matches with the relationship owner named, relevant upcoming events, and leaves a ready-to-send draft.

The human checkpoint is about 15 minutes each morning. Kill false positives, polish, forward. A weekly pass adds a hit-rate scorecard; a monthly pass looks 8 to 12 weeks ahead at the events calendar. A freshness check means a broken pipeline announces itself with an alert instead of failing silently.

Collection outside the AI, judgment inside it, data in the department's own system, and humans at exactly two checkpoints. No component depends on a personal laptop. Just as important: no single failure produces wrong output. Failures produce missing output, which is detectable, rather than confident nonsense that gets forwarded to senior agents.

1. Collect each watchlisted client's public activity.
2. Measure against that client's own 30-day baseline.
3. Flag spikes that cross the moment threshold.
4. Classify the moment type.
5. Match to the approved brand directory and events calendar.
6. Assemble a decision-ready brief with a draft.
7. Human review, about 15 minutes, then forward.

A freshness check alerts on a stale feed, so silence never masquerades as calm.

• A six-table data model as the system of record: a watchlist, a signals staging table, a metrics history table (the time series that makes per-client baselines possible), a brand directory, an events calendar, and the moment briefs queue.
• The real client roster loaded and structured, with social handles auto-extracted for the large majority and the rest flagged for manual fill.
• A daily automated analysis pass, prototyped end to end, that computes velocity against baseline, flags spikes, classifies moment type, matches against approved lists only, writes a brief, and prepares a draft.
• A brand-targeting taxonomy on four axes (audience, season and timing, vibe and context, and vertical fit) so a moment's tags filter the directory directly instead of relying on guesswork.
• A daily digest formatted for a 30-second phone scan: a one-line summary first, then moment blocks with only the decisive numbers, then a short watch list, then any data-quality flags, then only the items that actually need a human.
• A full architecture blueprint with a connector-by-connector evaluation separating what is load-bearing from what is expansion, plus the enterprise comparison to anchor a future budget ask.

Design complete, data store live, roster loaded, and the system shipped into a scoped 60-day pilot: a small cohort of clients, with the first two weeks reserved for baseline-building rather than detection.

I was deliberate about not overstating results. The pilot exists to produce hit-rate evidence; those numbers are its purpose, not a claim I get to make in advance. What I can claim is a working system, a governance-safe design, and a clear path to the evidence that justifies scaling it.

Collection was validated live at roughly ten cents per run, a few dollars a month at pilot scale, and a projected low tens of dollars a month at full roster. Buying collection and owning judgment beat building a scraper in-house against a moving target, at a cost that does not gate the decision.

I would run the pilot long enough to tune the moment threshold from a defensible default into a number derived from real hit-rate data. I would add per-client follower history to complement post-level signals. I would extend collection to a second platform once client authorization or a paid tool makes it clean rather than improvised. And I would build the two-sided version of the brand taxonomy so directory records carry the same audience and season tags the moments do, making the match symmetric.