Social Department replaces a week of binge-watching with four QuickSearch queries on AWS
A marketer doesn't need a frame with a coffee mug. They need the moment from a series that will spread across social media, and until now a show expert spent a week finding it. Social Department turned such questions into four QuickSearch queries on Amazon Bedrock: scenes, quotes, meme frames and trend matching. Video is tagged in advance and trends are collected separately. The two meet only at query time. That way an old series can be brought back into a current conversation without reprocessing.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
A marketer at a studio or streaming service rarely needs an object in the frame. They need a moment that will spread across social media: a scene, a line or a frame for a meme. Ordinary semantic video search is helpless here: it answers only the question you know how to phrase.
Ask for "a character with a coffee mug" and you get every mug from a thousand episodes. Which moment will resonate with the audience can't be typed into a search box. Until now, the answer came from a person who knew the show well and spent a week watching the material.
Social Department, which calls itself an AI operating system for entertainment marketing, turned such questions into four ready-made queries called QuickSearch. They run on Amazon Bedrock, the service where companies run models inside AWS infrastructure. The article was written by AWS engineers together with a Social Department representative.
TrackIt, an AWS partner, implemented the system based on the customer's architecture.
Four queries instead of a week of watching
Each QuickSearch query has its own search strategy, its own editorial criteria and its own ranking model. Shareable Scenes looks for complete scenes: prolonged conflicts, climaxes, staged episodes. Quotable Quotes picks short, quotable lines and surfaces those already circulating online. Memeable Moments selects a single frame and writes a caption based on what is happening in it. A separate scoring step chooses the frame, not a random thumbnail. Trend Align takes what people are discussing right now and finds matching moments in the library. A series that ended years ago can return to the live conversation. Plain text search and search by sample image remain available as well.
The company's principle is that AI proposes and people decide. Code checks the output against the source, and a quote that does not appear in the episode transcript is discarded. Bedrock Guardrails screens search queries for prompt attacks. A person on the client's team approves every post before it is published. Client data stays within AWS boundaries, is not passed to model developers and is not used to train their models.
How video processing works
The key design decision: video understanding and trend collection run independently, each on its own schedule. The library changes only when new episodes are added. Trends change hourly. So the two meet only at query time. Otherwise the whole catalog would have to be reprocessed for every fresh trend.
Uploading a clip to Amazon S3 triggers an AWS Step Functions chain of roughly two dozen Lambda functions. MediaConvert transcodes the video to 480p. Rekognition cuts it into 10–20 second ranges. The TwelveLabs Marengo 3.0 model builds embeddings from image, sound and speech, that is, numerical "fingerprints" of fragments for searching by meaning. TwelveLabs Pegasus 1.2 describes each scene: emotions, themes, dialogue type, place in the plot. Before that, a separate function cuts out the fragment with a margin of a few seconds at each edge so that transitions are not lost. Claude models build a character map. Everything is indexed in Amazon OpenSearch Serverless.
Each task retries with increasing backoff on failure, and scene description gets an additional retry for rate limiting. If an optional enrichment step fails, it does not take down the whole run. There is also a geographic caveat. Marengo is not available in every AWS region, so the video is temporarily placed in a bucket in the US East (N. Virginia) region. The rest of the infrastructure stays in the "home" region.
Trends: two channels and a filter
Trends are collected through two channels. The daily channel runs on an Amazon EventBridge schedule: one query per region per day, in the local morning. The "hot" channel, Heat Seeker, catches early topics that are gaining momentum. It fetches data only at the user's request and reuses what is already stored if it is fresh enough. This is a question of cost. The schedule pays for each region on every run, while on-demand fetching costs nothing on a quiet day.
There are four feeds: world, US, UK and Australia. This is not a filter over a shared list. Each region is collected separately, so trends on the same day really do differ between them. Raw trends are not shown to the client. Claude Haiku maps each trend to Social Department's own category schema, and then an editorial filter applies. The original text breaks off at this point, so we do not retell the details of the selection here.
What this means in practice
For anyone building video search or recommendations, the lesson is separation. It is better to process the library's content and a fast-changing external signal separately and join them at query time. That way, updating trends does not require reindexing. Subjective questions like "what will become a meme" should be turned into explicit editorial criteria, with results checked by code and by a person. The publication contains no quality metrics, such as the share of successful finds. It is a description of an architecture, not a benchmark, and the system cannot be compared with others on accuracy.
- Nov 19, 2026Amazon and Google will hold 28.4% of Anthropic after its IPO, but without voting rights
- Nov 19, 2026AWS AgentCore Gets Lambda Hooks and Uses 34% Fewer Tokens Than Alternatives
- Nov 19, 2026Skyflow in AgentCore Gateway hides personal data from AI agents' tools
- Nov 19, 2026Accenture Cuts Clients' Bedrock Token Costs by 40–85%
Key Questions Answered
What are the key details on Four queries instead of a week of watching?
Each QuickSearch query has its own search strategy, its own editorial criteria and its own ranking model. Shareable Scenes looks for complete scenes: prolonged conflicts, climaxes, staged episodes. Quotable Quotes picks short, quotable lines and surfaces those already circulating online. Memeable Moments selects a single frame and writes a caption based on what is happening in it. A separate scoring step chooses the frame, not a random thumbnail.
How video processing works?
The key design decision: video understanding and trend collection run independently, each on its own schedule. The library changes only when new episodes are added. Trends change hourly. So the two meet only at query time. Otherwise the whole catalog would have to be reprocessed for every fresh trend. Uploading a clip to Amazon S3 triggers an AWS Step Functions chain of roughly two dozen Lambda functions. MediaConvert transcodes the video to 480p. Rekognition cuts it into 10–20 second ranges.
What are the key details on Trends: two channels and a filter?
Trends are collected through two channels. The daily channel runs on an Amazon EventBridge schedule: one query per region per day, in the local morning. The "hot" channel, Heat Seeker, catches early topics that are gaining momentum. It fetches data only at the user's request and reuses what is already stored if it is fresh enough. This is a question of cost. The schedule pays for each region on every run, while on-demand fetching costs nothing on a quiet day. There are four feeds: world, US, UK and Australia.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.