Skip to main content
How It Works

From on-prem footage to a live FAST TV channel — how Kaster turns one video into a channel

Four stages, edge-first architecture — inventory & capture, AI cataloging & search, scheduling, and distribution — all handled automatically, with your original footage staying in your facility the entire time.

Kaster runs through four automated stages: inventory & capture, AI cataloging & search, scheduling, and distribution. The Edge Agent deploys on your existing NAS or an x86 machine (Docker); your archive is never moved or re-encoded. Only audio summaries, keyframes, and semantic embeddings — roughly 1–5% derivatives — are encrypted and sent to the cloud for analysis over an outbound-only HTTPS (443) connection, and cloud-side temporary data is auto-deleted under a 30-day lifecycle policy.

Four-Step Journey

The actual path a new video takes

01 — Inventory & Capture

Edge Agent scans your on-prem library

Automatically detects new NAS files, deduplicates via fingerprinting, and extracts keyframes and audio summaries — the original footage is never moved or re-encoded.

02 — AI Cataloging & Search

Semantic index and transcripts

A VLM builds a shot-level semantic index, and Whisper generates multilingual transcripts and captions — so a single sentence finds the exact shot in seconds.

03 — Scheduling & Production

Build a playlist and EPG

Pick your clips to build a channel playlist and set Break Rules; the system auto-generates the program schedule — no manual calendar work required.

04 — Publish & Distribute

Push to major platforms at once

The EPG and stream publish simultaneously to Samsung TV Plus, LG Channels, Pluto TV, Roku, and more, and the channel goes live.

Why not just compare similarity scores?
Vector similarity (cosine similarity) can't tell the difference between "this is in the shot" and "this isn't in the shot" — search for "on the phone" versus "taking off a jacket" and the similarity scores often land too close together to reliably separate what's actually relevant. Kaster instead scores each shot with a VLM across five separate dimensions — people, objects, scenes, actions, emotion — and only compares on the dimensions relevant to your query. The people dimension is cross-checked against a face database, so what comes back is an actual match, not just something that "looks kind of similar." See the full range of this multi-dimensional search in action in the AI semantic video search solution.

System & Hardware Architecture

What runs on-prem, what runs in the cloud, and how data actually moves

No GPU purchase, no new server room required. The Edge Agent installs directly on your existing NAS or a low-spec x86 machine (containerized via Docker), and only encrypts and sends 1–5% derivatives to the cloud — your archive stays on-prem the entire time.

On-Prem · Your Facility / NAS

HWHardware

Your existing NAS or a low-spec x86 machine, containerized via Docker — no GPU purchase, no new server room. Your current equipment is enough for continuous monitoring and sampled analysis.

EDGEEdge Agent

File monitoring and dedup (fingerprint matching) → keyframe extraction → audio summary (16kHz FLAC) → semantic embedding computation — all done on-prem.

EDGEOn-Prem Playout Engine

Scheduling logic and stream output run independently on-site with no cloud round-trip; even if the network drops, the channel keeps playing without a gap. See ad insertion details in the SSAI ad integration solution.

Cloud · Kaster AI

CLOUDIngest & Authorization

A Presigned URL, valid for a single use and expiring after 1 hour, with one key mapped to one purpose — only accepting derivatives actively uploaded by the Edge Agent.

CLOUDAI Analysis

VLM visual understanding, Whisper transcription, and vector indexing run in one pass — processing only the derivatives, and never seeing the complete original video.

CLOUDLifecycle Governance

Temporary analysis data is auto-deleted under a 30-day lifecycle policy — used once, then destroyed, with no residual copies retained.

Analysis results like indexes and captions sync back to the on-prem web console to power MAM search and scheduling; this return path also carries only derivatives and analysis results — the original video itself is never transmitted over this path.

1–3%
Audio summary (16kHz FLAC) as a share of the original file — enough for AI analysis, small enough to barely register
0.8%
Keyframe thumbnails (JPEG) as a share of the original file — enough visual signal without sending the whole video
0%
Share of original master footage uploaded to the cloud — not a single frame ever leaves your facility
Verified Support · Formats and Streaming Standards
Preview transcode
H.264 codec, MP4 container (faststart), for fast online preview
Ad-insertion stream
HLS, server-side insertion (SSAI) per SCTE-35 cue specification
For IT Teams

Architecture-level security specs: exposure risk eliminated at the source

Connection Direction (Outbound Only)
The Edge Agent only ever initiates outbound HTTPS (443) requests — your firewall never needs to open a single inbound port. External scanning is blocked at the physical network layer, closing off a malicious entry point and satisfying zero-trust architecture requirements.
Dynamic Derivative Transfer Authorization
Uploads use a Presigned URL mechanism — one key per single transfer, with a hard 60-minute expiration. Both the physical time limit and the narrow scope of use eliminate the risk of a leaked static link or unauthorized access.
Cloud Data Lifecycle Management
Once non-sensitive derivatives (audio summaries, keyframes, semantic embeddings) finish cloud analysis, they follow a strict 30-day auto-destruction lifecycle policy, ensuring the cloud environment never retains a redundant copy — a fully stateless design.
Control of Original Physical Files
Original high-resolution master footage is stored only, statically, on your own on-prem equipment (NAS/server); the system never backs it up, duplicates it, or uploads it to the cloud. Physical ownership and control of the data stays 100% inside your facility.
No-Trust-Boundary Design
Because the original footage never crosses your on-prem network boundary at all, data-leak risk is eliminated at the architecture-design stage, with no reliance on any additional trust mechanism or manual safeguard.
Access Permission Granularity
Permissions bind at three levels — user, role, and program (channel/content) — so authorization can be scoped by team function and content sensitivity, rather than a single all-or-nothing account.

Security is the first thing you should evaluate about Kaster — not a question you ask right before signing