Leader track 🎯 · six sessions, no code
For whoever runs the hosts and answers for the GMV. How to read a LIVE room's numbers, why the obvious playbook is a traffic effect, and how to write one that a host can follow and a test can confirm.
Read the LIVE room
The funnel from views to orders, the numbers TikTok Shop reports per stream, and the split that runs the whole course: peak viewers tells you about the traffic a room was given, GPM tells you how well it sold what it was given.
🏆 Leader session 2What the top streams share
Nine of the ten best nights opened with a giveaway, and inside every stream the giveaway minutes convert worst. The traffic effect that makes "top streams do X" the most expensive sentence in the business.
🗣️ Leader session 3Design the say
Hook, product build, price anchor, CTA and urgency as a script with measured minute values. The two-minute urgency cue that converts at 15 percent and the third minute that drops it to 8.
🎥 Leader session 4Design the show
What the camera adds on top of the words: product in hand, the close-up, the price on screen. Where a human host beats an AI avatar per viewer and where the avatar's cheap daytime hours still pay.
🎧 Leader session 5Coach the room live
A five-minute nowcast in the producer's ear, the three cues worth giving and the two the room begs for when it gets nervous. Run the replay bench yourself and see which ones the engine pays for.
📋 Leader session 6Prove the playbook worked
The scorecard a host and a script can be judged on, the test design that separates a cue from the traffic it rode, and the urgency claims the platform and the regulator both read as deception.
Builder track 🔧 · eight sessions, one pipeline
Sixty streams of Lumen Glow, one hero SKU, turned into a minute-level table and analysed for real in Python. Session 8 runs the finished playbook on tonight's stream and measures it over forty seeded nights.
The stream as data
Four sources, one grain: the video, the transcript, the comment feed and the per-minute ledger become one row per minute. The event model and the metric definitions everything downstream inherits.
⏱️ Builder session 2Ingest and align
FFmpeg segments, frames at 1 fps, 16 kHz audio, and the clock problem: a stream reaches viewers seconds after the host says it, so the sale lands in the wrong minute unless you shift it.
👁️ Builder session 3See it: what the camera shows
Product in hand, close-up ratio, host on camera, scene cuts and the price card read by OCR. Open-vocabulary detection against a named SKU, and the licence question that decides which detector you can ship.
🎙️ Builder session 4Hear it: what the host says
Word-timestamped transcripts, a segment tagger for hook, build, demo, price, CTA, urgency and Q&A, and a claims checker that flags the phrases a platform policy reads as deceptive.
📊 Builder session 5Score the content
Per-minute click rate, order rate and orders per thousand viewer-minutes by segment. The table where CTA minutes order at 11.6 per thousand and a hook orders at 2.0, and why a ratio on a small denominator lies.
🧭 Builder session 6From traits to a playbook
The pooled regression that prices a giveaway opener at plus 5,576 dollars, the within-stream version that prices the same minutes at minus 2.3 points of conversion, and a LightGBM with SHAP that agrees with the second.
📡 Builder session 7The live nowcast
Forecast the next five minutes of orders from the room so far. A smoothed order rate times a viewer projection, backtested over 6,720 windows, and the alert rule that does not fire on noise.
🎛️ Builder session 8Intervene, then prove it
The replay bench: cue the host at any minute of tonight's stream, run the same plan over forty nights, and read seven policies against the host's own. Then the switchback design that makes the result believable.
Choose your path 🗺️
Three doors into the same room. Both tracks start at their own session 1, because the funnel vocabulary is what makes the rest arguable.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Fourteen stamps across both tracks, and they live in this browser only.
What is real here 🧾
A course about separating a real effect from a traffic effect should be exact about which of its own numbers are measured, and which were planted so the methods could be checked against a known truth.
| Claim | Status |
|---|---|
| The 60 streams, 7,200 minute rows | Generated once by the course engine script from a fixed seed, identical for every learner and identical in Node and in your browser. Viewers arrive and leave, the pinned card is shown, clicks and orders are drawn minute by minute |
| The causal structure | Planted, on purpose, and documented in the engine. Which segments raise clicks, which raise orders, how a giveaway opener inflates early engagement and buys platform traffic while lowering conversion all night, how an urgency pitch stops converting from its third consecutive minute, how an AI host loses the room a little faster each time its script repeats. None of it is a claim about TikTok's real ranking. The point is that the builder-track methods can be checked against a truth that is known |
| Nine of the top ten opened with a giveaway; plus 5,576 dollars pooled, minus 2.3 points within-stream | Computed by the analysis script in the repository on the generated data: a pooled OLS across streams, then the same minutes demeaned by stream with errors clustered by stream. Both numbers are printed by the script |
| The segment table: CTA minutes order at 11.6 per thousand viewer-minutes, a hook at 1.8 | Arithmetic on the 7,200 rows, same script |
| The playbook: plus 27 percent a night; stretched urgency plus 0.5 percent and 3.2 policy-risk flags; a mid-stream giveaway minus 11 percent | Means over 40 seeded nights of tonight's stream, computed live by the bench in your browser and frozen by the same engine. A single night is shown too, and labelled as a single night |
| The nowcast beats "next five equals last five" | Backtested over 6,720 five-minute windows on the history, MAE 9.4 orders against 12.0. Printed by the analysis script |
| Tool versions and APIs (FFmpeg, PySceneDetect, Ultralytics YOLO26, MediaPipe, supervision, WhisperX, Gemini, LightGBM, SHAP, statsforecast) | Read from each project's docs or PyPI on 2026-09-24; versions are printed on the page that teaches them. Fast-moving: re-check before delivery |
| TikTok Shop LIVE metric names and policy wording | Read from TikTok Shop's own help and policy pages on 2026-09-24. Where a page could not be read the metric is taught by its arithmetic and marked so |
| Lumen Glow, its hosts, and TikTok's ranking rule | The brand and hosts are fictional. The boost rule is a teaching stand-in for "the platform pushes rooms that look engaged early", which is how the confound arises in any feed that rewards early engagement |
The knowledge map 🧠
Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.