IN-VIDEO AI
Video intelligence, as one API
Every video comes back as structured data: transcripts, chapters, entities, scenes and on-screen text. The same metadata powers clipping, search and moderation, and exports for your own RAG or recommendations.
Per-minute pricing, no seat count. $25 in signup credits, no credit card for the free tier.
TRUSTED BY PRODUCT TEAMS SHIPPING VIDEO AT SCALE
WHY IN-VIDEO AI
AI inside the encode, priced per minute
No second pipeline, no model to maintain, and the structured metadata is yours to export.
AI is done when the encode is done.
Flags go on the same upload call. No second pipeline, no second queue, no waiting on a job you have to poll.
You never own a model upgrade.
We pick, tune and re-benchmark the underlying models. Your request body does not change when they do.
One per-minute line item.
Priced beside encoding, per minute. No token billing and no separate AI invoice to forecast.
The metadata is yours.
Structured JSON on a webhook, exportable to your warehouse. Build the things we did not build.
Knovo runs AI video search across 200+ hours of course video. MyClassboard auto-chapters every lecture and feeds the transcript into completion events. Rocketlane uses it on in-product community video.
01 · WHAT COMES BACK
Turn any video into transcripts, chapters, and structured metadata
Set the flags on the upload call. When encoding finishes, a webhook delivers JSON that is already joined to one timeline, because it all came out of the same pass.
Video transcription with speaker diarization
transcript, speaker_labelsWord-level timestamps, speaker labels, 30+ languages. Read more
Video caption translations
translations[]Text translations of captions and transcripts. No synthetic audio. Read more
Auto video chapters
chapters[]Boundaries from visual and audio cues, each with a title. Read more
Video summarization
summary, topics[]Markdown, 100 to 300 words, with topic tags. Read more
Video entity recognition (NER)
entities[]People, orgs, places, products and dates, tied to the second they appeared. Read more
Video scene detection
scenes[]Shot change boundaries, so your chunks are not arbitrary.
Video object detection
objects[]Objects and labels per frame range, queryable without narration.
Video OCR (on-screen text)
ocr[]Slides, code, dashboards and lower thirds the speaker never reads out.
Advanced meeting summary
advanced_summary{}Overview, key decisions, action items, sentiment and speaker talk time as structured JSON. Read more
{
"asset_id": "a_9fK2",
"segments": [
{
"t_start": 862.4,
"t_end": 871.9,
"speaker": "S1",
"text": "retention drops right after the intro",
"chapter": "03 Pricing",
"entities": ["FastPix", "Q3"],
"objects": ["laptop", "whiteboard"],
"ocr": ["Q3 revenue"]
}
],
"advanced_summary": { "decisions": 3, "action_items": 5 }
}02 · READY TO USE
AI clipping, reframe, search, and moderation, ready to use
These are not a separate product. They read the same metadata you get back, which is the point: if clipping and search work on it, your own pipeline will too.
8 to 120 second clips
AI video clipping
Long video in, ranked short clips out. Highlight detection scores each on hook, pacing and narrative.
9:16 · 1:1 · 4:5 · 16:9
Video auto-reframe
16:9 to 9:16, 1:1 or 4:5, tracking the subject rather than centre-cropping. Built for Shorts, Reels and TikTok.
GET /search?q=
Video search
Conversational queries return the exact second a topic was mentioned, with the surrounding transcript context.
threshold tunable
Video moderation
NSFW and policy classification per frame, threshold tunable from 0.0 to 1.0, webhook on flag with an audit log. Read more
MEETING INTELLIGENCE
Advanced summaries with action items, decisions, and sentiment
Turn a webinar, interview or recording into structured, meeting-style notes. One request returns an overview, key decisions, action items with owners, sentiment and speaker talk time, every item timestamped. It is the same output the Notes Agent produces.
Overview and key decisions
meeting_summary, key_decisions[]A high-level snapshot plus the decisions made, each with an owner and a timestamp.
Action items
action_items[]Grouped by owner, with due dates when they are mentioned.
Sentiment and talk time
sentiment, speaker_talk_time[]Positive, neutral and negative split, plus how long each speaker held the floor.
Q3 Planning Review
45 minKEY DECISION
Approved budget for 2 engineers
15:32ACTION ITEMS
Share budget sheet
Jun 25Update the roadmap doc
Jul 02SENTIMENT
TALK TIME
03 · BUILD YOUR OWN
Build multimodal RAG, recommendations, and feed ranking
The same JSON goes to your warehouse. What you do with it is not something we need to have shipped first.
chunk → embed → retrieve
Multimodal RAG over video
Timestamped transcript, chapters, OCR and detections are the extraction and chunking layer of a RAG pipeline. Chapter boundaries give you semantically whole chunks instead of fixed 30-second windows that cut mid-sentence.
content features + engagement
Recommendations
Content features from In-Video AI plus engagement from Video Data, landing in the same warehouse. Those are the two inputs a content-based recommender needs, and most vendors give you only one of them.
clip-level signals per unit
Feed ranking
Clip-level metadata means every short-form unit is described individually, not inherited from its parent asset. Rank a vertical feed on what is actually in each clip.
What you still own
Embeddings, the vector store, the retrieval logic and the model. FastPix gives you the extraction and chunking layer, not a vector database. If you want the embeddings themselves under your control, that is the right way round: you are not locked into ours.
BUILD VS BUY
The same video AI, assembled on AWS or Google Cloud
Most teams evaluating this are choosing between one API and a stack they wire together. Here is the honest shape of that choice.
| Capability | On AWS | On Google Cloud | On FastPix |
|---|---|---|---|
| Transcript with speaker labels | Transcribe (+ diarization) | Speech-to-Text (+ diarization) | One flag |
| Named entities | Comprehend | Natural Language API | One flag |
| On-screen text (OCR) | Rekognition text detection | Video Intelligence text detection | One flag |
| Scene and object detection | Rekognition Video segments + labels | Video Intelligence shot change + labels | One flag |
| Chapters and summary | Bedrock | Gemini on Vertex AI | One flag |
| Semantic search | OpenSearch + embeddings | Vertex AI Vector Search | One flag |
| Clipping and reframe | MediaConvert | Transcoder API | One flag |
| Moderation | Rekognition content moderation | Video Intelligence explicit content | One flag |
| Orchestration | Step Functions + Lambda | Cloud Workflows | One flag |
The glue is the project, not the API calls.
Seven or more services, each with its own retry semantics, quota, failure mode and per-asset cost line. The integration work is orchestration, partial-failure handling and cost accounting.
Every AI step after encoding re-reads the mezzanine.
Run AI as a second pass and you pull the source out of storage again for each service. FastPix runs AI inside the encode pass, so the file is decoded once.
Independently-run services return different clocks.
Joining transcript words to shot boundaries to OCR frames onto one timeline is where teams lose weeks. The output arrives pre-joined here because it came from one pass.
When building it yourself is the right call
You need a domain-tuned model. Medical, legal or heavily accented ASR often beats a general model, and fine-tuning is not something you can flag on an upload call.
You have committed cloud spend to burn. If you are already drawing down an AWS or Google commitment, the marginal cost of those services is not the list price.
You need the embeddings themselves under your control, in your own vector store, with your own model. Our search endpoint is an endpoint, not a vector database.
CAPABILITIES THAT SHIP
Every video AI capability, behind one flag
Tech specs
What In-Video AI handles.
Features, languages, output formats, integration patterns.
Transcription languages
Caption formats
Chapter detection
Search
Summary
Moderation
NER
Integration
Questions developers ask
In-Video AI questions, answered.
What is a video understanding API?
It turns a video file into structured data you can query. FastPix returns transcripts with speaker diarization, auto chapters, named entities, scene and object detection, and on-screen text as JSON on a webhook. You set flags on the upload call rather than running a second pipeline after encoding.
Can I build RAG over video with this?
Yes, for the extraction and chunking half of the pipeline. Timestamped diarized transcripts, chapter boundaries, OCR and detections are exactly what a retrieval layer ingests, and chapters give you chunk boundaries that do not cut mid-sentence. You still own the embeddings, the vector store and the model.
How is this different from running my own model?
You do not pick or maintain the model. Set a flag at upload and the output is part of the asset. We tune and re-benchmark the underlying models, and billing stays per minute alongside encoding.
What is an advanced summary?
A structured, meeting-style summary of a recording. One request returns an overview, key decisions, action items with owners, sentiment and speaker talk time as JSON, every item timestamped. The Notes Agent produces the same output automatically.
Pricing
Per-minute AI processing.
Inline with encoding. Same per-minute billing model. See full pricing.
TRANSCRIPTION + CAPTIONS
Per minute transcribed.
$0.048/ minute
30+ languages with VTT and SRT export. Same rate for live auto-generated subtitles.
- 30+ languages
- VTT + SRT export
- Speaker diarization
SEARCH + SUMMARY + NER
Per minute analyzed.
$0.0035/ minute
Markdown summaries, structured entities, video chapters, and a conversational search endpoint. Same per-minute rate across NER, chapters, and summary.
- Conversational search endpoint
- Markdown summary
- Named-entity recognition + chapters
MODERATION + SCENE DETECTION
Per minute moderated.
$0.10/ minute
Tunable NSFW / profanity classifier with audit-log-grade webhooks.
- NSFW + profanity detection
- Threshold tuneable per workload
- Webhook + audit log per decision
Three ways to get unstuck
Whatever kind of help you need, there is a path.
Engineering support
Talk to a video engineer.
Stuck on an API call, a webhook signature, or a player integration? Reach the engineering team directly. Response within hours, not days.
Contact engineeringIntegration help
Docs, code samples, video tutorials.
Self-serve resources for the most common integrations. Quickstart guides, SDK examples, and detailed playback logs in your dashboard.
Browse the docsSolution architect
Plan the rollout with a human.
New integration, migration off another platform, or a complex multi-tenant build. Book a session with a FastPix solution architect.
Join the Slack communityDeveloper resources
Everything you need to start building.
Five-minute quick-start
Sign up, hit the endpoint, ship.
Quick-start guideFull API reference
Every endpoint, every parameter, every response.
API referenceWebhook reference
Every event FastPix emits, with sample payloads.
WebhooksCode samples
Sample apps and SDK examples on GitHub.
GitHubSlack community
Talk to FastPix engineers and other developers.
Join SlackService status
Real-time uptime and incident reports.
Status page