K-OTT — Cross-OTT Recommendation Engine

At a Glance

6

OTT Platforms

~425

Rows Indexed

10

AI Agents

<1s

First Response

Why K-OTT

Korea's OTT market is fragmented across many platforms. Users juggle multiple subscriptions, each with different content libraries, recommendation algorithms, and UI paradigms. Finding what to watch tonight shouldn't be a 30-minute ordeal.

K-OTT solves this by integrating availability across 6 Korean OTT platforms — Netflix, TVING, Coupang Play, Wavve, Disney+, and Watcha — into a single AI-powered interface. Ask in natural language, and the engine parses your intent (mood, genre, exclusions) to recommend what's actually available right now.

How it works

User Query — IntentParser (OTT/genre/negation extraction)
   — MOOD_TO_GENRES (sentiment — genre mapping)
     — Hybrid Search (PostgreSQL FTS + pgvector RRF)
       — RecommendEngine — Gemini API
         — SSE 5-chunk streaming:
          thought — metadata — message — xai — actions (deep links)

Triple Guardrail

Layer 1: Availability — Cross-checks OTT platform availability in real-time
Layer 2: Relevance — Filters by genre, mood, and user-specified exclusions
Layer 3: Deduplication — Prevents repeated recommendations across sessions

AI Agents (10)

Agent	Role	Cycle
subscription_optimizer	Optimal OTT combo based on viewing patterns	Weekly
content_crawler	New content collection across 7 platforms	Daily
review_analyzer	Sentiment analysis + keyword extraction	Daily
pipeline_resilience	Failure detection — auto-retry — Telegram alert	Always-on
youtube_curator	Trailer matching for recommended content	Daily
recommendation	Recommendation orchestration	Real-time
youtube_quota_manager	YouTube API quota optimization (10K units/day)	Always-on
trend_ranker	SNS buzz + search volume + TMDB popularity	Daily
orchestration	Agent execution order coordination	Triggered
intent_parser	User intent analysis (OTT/genre/negation)	Real-time

Tech Stack

FastAPI Next.js Supabase pgvector Gemini AI TMDB API Redis Vercel