From f00fc73b0bfc4f7c11722c5b901bedd61783cb42 Mon Sep 17 00:00:00 2001
From: ayrisdev <ayris.dev@gmail.com>
Date: Tue, 1 Sep 2026 15:26:22 +0300
Subject: [PATCH] =?UTF-8?q?feat:=209:16=20crop=20+=20dinamik=20altyaz?=
 =?UTF-8?q?=C4=B1=20render=20motoru=20(Faz=202,=20Mod=C3=BCl=204)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Her aday klip transkribe olduktan hemen sonra otomatik olarak
video-render kuyruğuna giriyor. Yeni Python worker (video_render.py):

- Klibi ffmpeg ile keser
- OpenCV + MediaPipe Face Detector (Tasks API — yeni mediapipe
  sürümlerinde eski solutions API'si kalktı, .tflite model dosyası
  Docker build'de indiriliyor) ile örneklenmiş karelerde yüz merkezi
  tespit eder, yumuşatır (moving average); hiç yüz bulunamazsa sabit
  merkez crop'a düşer, hata saymaz
- Kırpılmış kareleri OpenCV'den ffmpeg'e pipe'layıp tek geçişte orijinal
  sesi map'ler ve transcript_json'dan ürettiği .ass altyazıyı yakar
- Sonucu /shared-media/shorts/<candidate_id>.mp4'e yazar

Yeni ShortVideo modeli (PENDING/RENDERING/READY/FAILED). Panelde
(segments + kanal detay sayfaları) render durumu + READY olunca video
önizleme/indirme. VIDEO_RENDER_CONCURRENCY ayrı ve düşük (varsayılan 2)
— bu iş diğerlerinden çok daha ağır.

Kapsam: sadece tekil konuşmacı şablonu (çoklu konuşmacı/split-screen
sonraki bir iterasyona bırakıldı — aktif-konuşmacı tespiti gerektiriyor).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---
 .env.example                                  |   5 +
 README.md                                     |  39 ++-
 .../app/api/segments/[id]/video/route.ts      |  40 +--
 .../app/api/shorts/[id]/video/route.ts        |  12 +
 apps/frontend/app/channels/[id]/page.tsx      |  26 +-
 apps/frontend/app/segments/page.tsx           |  26 +-
 apps/frontend/lib/videoStream.ts              |  42 +++
 apps/worker/Dockerfile                        |   8 +-
 apps/worker/requirements.txt                  |   2 +
 apps/worker/worker/db.py                      |  58 ++++
 apps/worker/worker/main.py                    |   5 +-
 apps/worker/worker/queues.py                  |   5 +
 apps/worker/worker/stt_scoring.py             |   5 +-
 apps/worker/worker/video_render.py            | 260 ++++++++++++++++++
 docker-compose.yml                            |   1 +
 .../20260901010000_short_videos/migration.sql |  21 ++
 packages/db/prisma/schema.prisma              |  22 ++
 17 files changed, 519 insertions(+), 58 deletions(-)
 create mode 100644 apps/frontend/app/api/shorts/[id]/video/route.ts
 create mode 100644 apps/frontend/lib/videoStream.ts
 create mode 100644 apps/worker/worker/video_render.py
 create mode 100644 packages/db/prisma/migrations/20260901010000_short_videos/migration.sql

diff --git a/.env.example b/.env.example
index 94c4415..cf958ec 100644
--- a/.env.example
+++ b/.env.example
@@ -50,3 +50,8 @@ PROXY_URL=
 # (DB row + file) — PRD's auto-purge lifecycle. Only segments already fully
 # analyzed (no candidate still waiting on STT) are eligible.
 RAW_SEGMENT_TTL_HOURS=24
+
+# Video render (yüz takibi + 9:16 crop + altyazı) her segment/stt-scoring
+# işinden çok daha ağır — düşük tutulur ki aynı anda birden fazla render
+# worker container'ını boğmasın.
+VIDEO_RENDER_CONCURRENCY=2
diff --git a/README.md b/README.md
index 98fcd38..2db5e12 100644
--- a/README.md
+++ b/README.md
@@ -1,11 +1,14 @@
-# StreamClipper AI — Faz 1
+# StreamClipper AI — Faz 1 + Faz 2 (render)
 
-`prd.md` içindeki tam kapsamlı otonom pipeline'ın ilk fazı: ingestion (yt-dlp
-headless capture + 15dk segmentleme), sinyal analizi (ses peak + chat
-velocity), Whisper STT (OpenAI API, kelime zaman damgalı), BullMQ/Redis/
-Postgres altyapısı ve durum/transkript gösteren bir Next.js panel. LLM
-virality skorlama, render/crop/altyazı ve multi-platform dağıtım bu fazda
-**yok** — bkz. `prd.md` Modül 3 (skorlama kısmı), 4 ve 5.
+`prd.md` içindeki tam kapsamlı otonom pipeline'ın Faz 1+2'si: ingestion
+(yt-dlp headless capture + 15dk segmentleme), sinyal analizi (ses peak +
+chat velocity), Whisper STT (OpenAI API, kelime zaman damgalı), BullMQ/
+Redis/Postgres altyapısı, auth korumalı bir Next.js panel — ve Faz 2 olarak
+transkribe edilen her aday klibi otomatik yüz-takipli 9:16'ya crop'layıp
+dinamik altyazı basan bir render motoru (`apps/worker/worker/video_render.py`).
+LLM virality skorlama, çoklu-konuşmacı/split-screen şablonu ve
+multi-platform dağıtım hâlâ **yok** — bkz. `prd.md` Modül 3 (skorlama
+kısmı) ve Modül 5.
 
 ## Klasör Yapısı
 
@@ -77,8 +80,8 @@ Panel: http://localhost:3000 · API daemon health: http://localhost:4001/health
 Panelin tüm sayfaları giriş gerektirir (middleware ile korunuyor). İlk açılışta
 henüz hiç kullanıcı yoksa `/login` otomatik olarak "İlk Yönetici Hesabını
 Oluştur" formunu gösterir — kullanıcı adı/şifre girip ilk admin hesabını
-oradan oluşturursun. Sonraki kullanıcılar için ayrı bir "kullanıcı ekle"
-arayüzü henüz yok (Faz 2).
+oradan oluşturursun. `/users` sayfasından ek kullanıcı ekleyip silebilirsin
+(kendi hesabını veya son kalan tek kullanıcıyı silmek engellenir).
 
 `SESSION_SECRET` production'da zorunlu (`openssl rand -base64 32` ile
 üretilebilir) — local dev'de boş bırakılırsa güvensiz bir varsayılan
@@ -94,15 +97,21 @@ kullanılır, sadece test için.
    kuyruğundan alıp ses peak + (varsa) chat velocity analiziyle
    `candidate_segments` üretir.
 4. Her aday, `stt-scoring` kuyruğunda OpenAI Whisper API ile transkribe edilir;
-   sonuç `candidate_segments.transcript_json` alanına yazılır ve Telegram'a
-   bildirim gider.
-5. Panelde (`/` ve `/segments`) kanal durumu ve üretilen transkriptleri
-   kontrol et.
+   sonuç `candidate_segments.transcript_json` alanına yazılır, Telegram'a
+   bildirim gider ve `video-render` kuyruğuna otomatik iş eklenir.
+5. `video-render` (Python, `video_render.py`) yüz tespiti + yumuşatılmış
+   dinamik crop ile 9:16'ya çevirir, transkriptten `.ass` altyazı üretip
+   yakar, `/shared-media/shorts/<candidate_id>.mp4`'e yazar;
+   `short_videos.status` `READY`/`FAILED` olur, Telegram'a bildirim gider.
+6. Panelde (`/`, `/segments`, kanal detay sayfaları) kanal durumu, ham
+   segment/transkript ve render edilen 9:16 kısa videoları izle/indir.
 
 ## Bilinen Sınırlamalar (bu faz kapsamı dışı)
 
-- LLM virality skorlama, metadata üretimi, telif/görüntü filtresi yok.
-- 9:16 crop/altyazı render motoru yok.
+- LLM virality skorlama, metadata üretimi, telif/görüntü filtresi yok —
+  transkribe edilen her aday otomatik render edilir, eşik/filtre yok.
+- Çoklu konuşmacı / split-screen şablonu yok — sadece tekil konuşmacı
+  (yüz-merkezli dinamik crop).
 - YouTube/Instagram/TikTok'a otomatik yayınlama yok (bu API'ler ayrı onay
   süreçleri gerektiriyor — `prd.md` tartışmasına bkz.).
 - Aday pencere çıkarımı her zaman tek bir 15dk'lık segment dosyasıyla sınırlı;
diff --git a/apps/frontend/app/api/segments/[id]/video/route.ts b/apps/frontend/app/api/segments/[id]/video/route.ts
index 7c00eb3..eecf4a5 100644
--- a/apps/frontend/app/api/segments/[id]/video/route.ts
+++ b/apps/frontend/app/api/segments/[id]/video/route.ts
@@ -1,7 +1,6 @@
-import { createReadStream, statSync } from "node:fs";
-import { Readable } from "node:stream";
 import type { NextRequest } from "next/server";
 import { prisma } from "@streamclipper/db";
+import { streamVideoFile } from "../../../../../lib/videoStream";
 
 export async function GET(req: NextRequest, { params }: { params: Promise<{ id: string }> }) {
   const { id } = await params;
@@ -9,40 +8,5 @@ export async function GET(req: NextRequest, { params }: { params: Promise<{ id:
   const segment = await prisma.rawSegment.findUnique({ where: { id } });
   if (!segment) return new Response("Not found", { status: 404 });
 
-  let size: number;
-  try {
-    size = statSync(segment.filePath).size;
-  } catch {
-    return new Response("File not found on disk", { status: 404 });
-  }
-
-  const headers: Record<string, string> = {
-    "Content-Type": "video/mp4",
-    "Accept-Ranges": "bytes",
-  };
-  if (req.nextUrl.searchParams.get("download") === "1") {
-    headers["Content-Disposition"] = `attachment; filename="${id}.mp4"`;
-  }
-
-  const range = req.headers.get("range");
-  const match = range ? /bytes=(\d+)-(\d*)/.exec(range) : null;
-
-  if (match) {
-    const start = Number(match[1]);
-    const end = match[2] ? Number(match[2]) : size - 1;
-    const stream = createReadStream(segment.filePath, { start, end });
-    return new Response(Readable.toWeb(stream) as ReadableStream, {
-      status: 206,
-      headers: {
-        ...headers,
-        "Content-Range": `bytes ${start}-${end}/${size}`,
-        "Content-Length": String(end - start + 1),
-      },
-    });
-  }
-
-  const stream = createReadStream(segment.filePath);
-  return new Response(Readable.toWeb(stream) as ReadableStream, {
-    headers: { ...headers, "Content-Length": String(size) },
-  });
+  return streamVideoFile(req, segment.filePath, `${id}.mp4`);
 }
diff --git a/apps/frontend/app/api/shorts/[id]/video/route.ts b/apps/frontend/app/api/shorts/[id]/video/route.ts
new file mode 100644
index 0000000..4371dff
--- /dev/null
+++ b/apps/frontend/app/api/shorts/[id]/video/route.ts
@@ -0,0 +1,12 @@
+import type { NextRequest } from "next/server";
+import { prisma } from "@streamclipper/db";
+import { streamVideoFile } from "../../../../../lib/videoStream";
+
+export async function GET(req: NextRequest, { params }: { params: Promise<{ id: string }> }) {
+  const { id } = await params;
+
+  const short = await prisma.shortVideo.findUnique({ where: { id } });
+  if (!short?.filePath) return new Response("Not found", { status: 404 });
+
+  return streamVideoFile(req, short.filePath, `${id}.mp4`);
+}
diff --git a/apps/frontend/app/channels/[id]/page.tsx b/apps/frontend/app/channels/[id]/page.tsx
index 09b0970..3b14a35 100644
--- a/apps/frontend/app/channels/[id]/page.tsx
+++ b/apps/frontend/app/channels/[id]/page.tsx
@@ -13,6 +13,8 @@ const STATUS_BADGE: Record<string, string> = {
   DISCARDED: "err",
   PENDING_STT: "warn",
   TRANSCRIBED: "ok",
+  RENDERING: "warn",
+  READY: "ok",
   FAILED: "err",
 };
 
@@ -34,7 +36,7 @@ export default async function ChannelDetailPage({ params }: { params: Promise<{
           include: {
             segments: {
               orderBy: { createdAt: "desc" },
-              include: { candidates: { orderBy: { createdAt: "desc" } } },
+              include: { candidates: { orderBy: { createdAt: "desc" }, include: { short: true } } },
             },
           },
         },
@@ -132,6 +134,28 @@ export default async function ChannelDetailPage({ params }: { params: Promise<{
                       {transcriptPreview(c.transcriptJson) && (
                         <p className="transcript-preview">{transcriptPreview(c.transcriptJson)}</p>
                       )}
+
+                      {c.short && (
+                        <div style={{ marginTop: "0.5rem" }}>
+                          <span className={`badge ${STATUS_BADGE[c.short.status] ?? "muted"}`}>
+                            9:16: {c.short.status}
+                          </span>
+                          {c.short.status === "READY" && (
+                            <>
+                              <video controls preload="metadata" style={{ width: "220px", marginTop: "0.4rem", borderRadius: "6px", display: "block" }}>
+                                <source src={`/api/shorts/${c.short.id}/video`} type="video/mp4" />
+                              </video>
+                              <a className="badge muted" style={{ marginTop: "0.3rem", display: "inline-block" }} href={`/api/shorts/${c.short.id}/video?download=1`}>
+                                İndir
+                              </a>
+                            </>
+                          )}
+                          {c.short.status === "FAILED" && c.short.errorMessage && (
+                            <p className="mono" style={{ fontSize: "0.75rem", marginTop: "0.3rem" }}>{c.short.errorMessage}</p>
+                          )}
+                        </div>
+                      )}
+
                       <form action={deleteCandidateSegment.bind(null, c.id)} style={{ marginTop: "0.35rem" }}>
                         <DeleteButton confirmText="Bu aday klibi silmek istediğine emin misin?" />
                       </form>
diff --git a/apps/frontend/app/segments/page.tsx b/apps/frontend/app/segments/page.tsx
index 012efa1..725d19d 100644
--- a/apps/frontend/app/segments/page.tsx
+++ b/apps/frontend/app/segments/page.tsx
@@ -10,6 +10,8 @@ const STATUS_BADGE: Record<string, string> = {
   DISCARDED: "err",
   PENDING_STT: "warn",
   TRANSCRIBED: "ok",
+  RENDERING: "warn",
+  READY: "ok",
   FAILED: "err",
 };
 
@@ -23,7 +25,7 @@ export default async function SegmentsPage() {
   const segments = await prisma.rawSegment.findMany({
     orderBy: { createdAt: "desc" },
     take: 50,
-    include: { candidates: { orderBy: { createdAt: "desc" } } },
+    include: { candidates: { orderBy: { createdAt: "desc" }, include: { short: true } } },
   });
 
   return (
@@ -71,6 +73,28 @@ export default async function SegmentsPage() {
                   {transcriptPreview(c.transcriptJson) && (
                     <p className="transcript-preview">{transcriptPreview(c.transcriptJson)}</p>
                   )}
+
+                  {c.short && (
+                    <div style={{ marginTop: "0.5rem" }}>
+                      <span className={`badge ${STATUS_BADGE[c.short.status] ?? "muted"}`}>
+                        9:16: {c.short.status}
+                      </span>
+                      {c.short.status === "READY" && (
+                        <>
+                          <video controls preload="metadata" style={{ width: "220px", marginTop: "0.4rem", borderRadius: "6px", display: "block" }}>
+                            <source src={`/api/shorts/${c.short.id}/video`} type="video/mp4" />
+                          </video>
+                          <a className="badge muted" style={{ marginTop: "0.3rem", display: "inline-block" }} href={`/api/shorts/${c.short.id}/video?download=1`}>
+                            İndir
+                          </a>
+                        </>
+                      )}
+                      {c.short.status === "FAILED" && c.short.errorMessage && (
+                        <p className="mono" style={{ fontSize: "0.75rem", marginTop: "0.3rem" }}>{c.short.errorMessage}</p>
+                      )}
+                    </div>
+                  )}
+
                   <form action={deleteCandidateSegment.bind(null, c.id)} style={{ marginTop: "0.35rem" }}>
                     <DeleteButton confirmText="Bu aday klibi silmek istediğine emin misin?" />
                   </form>
diff --git a/apps/frontend/lib/videoStream.ts b/apps/frontend/lib/videoStream.ts
new file mode 100644
index 0000000..d952291
--- /dev/null
+++ b/apps/frontend/lib/videoStream.ts
@@ -0,0 +1,42 @@
+import { createReadStream, statSync } from "node:fs";
+import { Readable } from "node:stream";
+import type { NextRequest } from "next/server";
+
+export function streamVideoFile(req: NextRequest, filePath: string, downloadName: string): Response {
+  let size: number;
+  try {
+    size = statSync(filePath).size;
+  } catch {
+    return new Response("File not found on disk", { status: 404 });
+  }
+
+  const headers: Record<string, string> = {
+    "Content-Type": "video/mp4",
+    "Accept-Ranges": "bytes",
+  };
+  if (req.nextUrl.searchParams.get("download") === "1") {
+    headers["Content-Disposition"] = `attachment; filename="${downloadName}"`;
+  }
+
+  const range = req.headers.get("range");
+  const match = range ? /bytes=(\d+)-(\d*)/.exec(range) : null;
+
+  if (match) {
+    const start = Number(match[1]);
+    const end = match[2] ? Number(match[2]) : size - 1;
+    const stream = createReadStream(filePath, { start, end });
+    return new Response(Readable.toWeb(stream) as ReadableStream, {
+      status: 206,
+      headers: {
+        ...headers,
+        "Content-Range": `bytes ${start}-${end}/${size}`,
+        "Content-Length": String(end - start + 1),
+      },
+    });
+  }
+
+  const stream = createReadStream(filePath);
+  return new Response(Readable.toWeb(stream) as ReadableStream, {
+    headers: { ...headers, "Content-Length": String(size) },
+  });
+}
diff --git a/apps/worker/Dockerfile b/apps/worker/Dockerfile
index 06bc407..bace075 100644
--- a/apps/worker/Dockerfile
+++ b/apps/worker/Dockerfile
@@ -1,7 +1,7 @@
 FROM python:3.12-slim
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
-      ffmpeg libsndfile1 \
+      ffmpeg libsndfile1 libgl1 libglib2.0-0 curl \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
@@ -9,5 +9,11 @@ COPY apps/worker/requirements.txt ./apps/worker/requirements.txt
 RUN pip install --no-cache-dir -r apps/worker/requirements.txt
 COPY apps/worker ./apps/worker
 
+# MediaPipe's Tasks API needs the face detector model asset on disk — not
+# bundled with the pip package, downloaded once at build time instead of on
+# every worker startup.
+RUN mkdir -p /app/models && curl -L -o /app/models/blaze_face_short_range.tflite \
+      https://storage.googleapis.com/mediapipe-models/face_detector/blaze_face_short_range/float16/latest/blaze_face_short_range.tflite
+
 WORKDIR /app/apps/worker
 CMD ["python", "-m", "worker.main"]
diff --git a/apps/worker/requirements.txt b/apps/worker/requirements.txt
index d4b20bd..91dcf1e 100644
--- a/apps/worker/requirements.txt
+++ b/apps/worker/requirements.txt
@@ -6,3 +6,5 @@ openai>=1.50.0
 httpx>=0.27.0
 chat-downloader>=0.2.7
 python-dotenv>=1.0.0
+opencv-python-headless>=4.10.0
+mediapipe>=0.10.14
diff --git a/apps/worker/worker/db.py b/apps/worker/worker/db.py
index 41912a2..817c3f6 100644
--- a/apps/worker/worker/db.py
+++ b/apps/worker/worker/db.py
@@ -93,3 +93,61 @@ async def update_candidate_transcript(candidate_id: str, transcript: dict, statu
 async def get_active_sessions() -> list[asyncpg.Record]:
     pool = await get_pool()
     return await pool.fetch("SELECT id, live_video_id FROM stream_sessions WHERE ended_at IS NULL")
+
+
+async def get_candidate_for_render(candidate_id: str) -> asyncpg.Record | None:
+    pool = await get_pool()
+    return await pool.fetchrow(
+        """
+        SELECT cs.id, cs.start_sec, cs.end_sec, cs.transcript_json, rs.file_path, c.name AS channel_name
+        FROM candidate_segments cs
+        JOIN raw_segments rs ON rs.id = cs.raw_segment_id
+        JOIN stream_sessions ss ON ss.id = rs.session_id
+        JOIN channels c ON c.id = ss.channel_id
+        WHERE cs.id = $1
+        """,
+        candidate_id,
+    )
+
+
+async def upsert_short_video(candidate_id: str, status: str) -> str:
+    pool = await get_pool()
+    existing = await pool.fetchrow(
+        "SELECT id FROM short_videos WHERE candidate_id = $1", candidate_id
+    )
+    if existing:
+        await pool.execute(
+            "UPDATE short_videos SET status = $2, error_message = NULL, updated_at = now() WHERE id = $1",
+            existing["id"],
+            status,
+        )
+        return existing["id"]
+
+    short_id = str(uuid.uuid4())
+    await pool.execute(
+        """
+        INSERT INTO short_videos (id, candidate_id, status, created_at, updated_at)
+        VALUES ($1, $2, $3, now(), now())
+        """,
+        short_id,
+        candidate_id,
+        status,
+    )
+    return short_id
+
+
+async def update_short_video(
+    short_id: str, status: str, file_path: str | None = None, error_message: str | None = None
+) -> None:
+    pool = await get_pool()
+    await pool.execute(
+        """
+        UPDATE short_videos
+        SET status = $2, file_path = COALESCE($3, file_path), error_message = $4, updated_at = now()
+        WHERE id = $1
+        """,
+        short_id,
+        status,
+        file_path,
+        error_message,
+    )
diff --git a/apps/worker/worker/main.py b/apps/worker/worker/main.py
index 32e7e70..e7e7577 100644
--- a/apps/worker/worker/main.py
+++ b/apps/worker/worker/main.py
@@ -4,6 +4,7 @@ from . import db
 from .chat_capture import start_chat_capture
 from .signal_detection import start_signal_detection_worker
 from .stt_scoring import start_stt_scoring_worker
+from .video_render import start_video_render_worker
 
 SESSION_WATCH_INTERVAL_SEC = 10
 
@@ -22,13 +23,15 @@ async def watch_sessions_for_chat_capture():
 async def main():
     signal_worker = start_signal_detection_worker()
     stt_worker = start_stt_scoring_worker()
-    print("[main] worker started: signal-detection + stt-scoring queues")
+    render_worker = start_video_render_worker()
+    print("[main] worker started: signal-detection + stt-scoring + video-render queues")
 
     try:
         await watch_sessions_for_chat_capture()
     finally:
         await signal_worker.close()
         await stt_worker.close()
+        await render_worker.close()
 
 
 if __name__ == "__main__":
diff --git a/apps/worker/worker/queues.py b/apps/worker/worker/queues.py
index b7599ea..4d1d5c1 100644
--- a/apps/worker/worker/queues.py
+++ b/apps/worker/worker/queues.py
@@ -21,3 +21,8 @@ SHARED_MEDIA_ROOT = os.environ.get("SHARED_MEDIA_ROOT", "./shared-media")
 # producing segments/candidates around the same time would otherwise process
 # strictly one at a time and build up a backlog.
 WORKER_CONCURRENCY = int(os.environ.get("WORKER_CONCURRENCY", "5"))
+
+# Video rendering (face detection across sampled frames + two ffmpeg encodes)
+# is far heavier per job than signal-detection/stt-scoring — kept separate
+# and lower by default so several renders at once don't starve the container.
+VIDEO_RENDER_CONCURRENCY = int(os.environ.get("VIDEO_RENDER_CONCURRENCY", "2"))
diff --git a/apps/worker/worker/stt_scoring.py b/apps/worker/worker/stt_scoring.py
index c429a5c..9fc132f 100644
--- a/apps/worker/worker/stt_scoring.py
+++ b/apps/worker/worker/stt_scoring.py
@@ -3,7 +3,7 @@ import os
 import subprocess
 import tempfile
 
-from bullmq import Worker
+from bullmq import Queue, Worker
 from openai import OpenAI
 
 from . import db
@@ -11,6 +11,7 @@ from .queues import QUEUE_NAMES, REDIS_URL, WORKER_CONCURRENCY
 from .telegram import send_telegram_message
 
 _openai_client: OpenAI | None = None
+video_render_queue = Queue(QUEUE_NAMES["VIDEO_RENDER"], {"connection": REDIS_URL})
 
 
 def _client() -> OpenAI:
@@ -67,6 +68,8 @@ async def process_stt_scoring(job, job_token=None):
         f"({candidate['start_sec']}-{candidate['end_sec']}s):\n_{preview}_"
     )
 
+    await video_render_queue.add("render-short", {"candidateSegmentId": candidate_id})
+
     print(f"[stt-scoring] candidate {candidate_id} transcribed")
 
 
diff --git a/apps/worker/worker/video_render.py b/apps/worker/worker/video_render.py
new file mode 100644
index 0000000..d86f173
--- /dev/null
+++ b/apps/worker/worker/video_render.py
@@ -0,0 +1,260 @@
+"""
+9:16 crop (yüz takipli) + dinamik altyazı render motoru.
+
+İki geçişli işleme: (1) OpenCV ile örneklenmiş karelerde MediaPipe yüz
+tespiti yapıp bir merkez-yol haritası çıkarır, yumuşatır; (2) OpenCV ile
+kareleri tekrar okuyup o an için hesaplanan crop penceresini uygulayıp
+ffmpeg'e pipe'lar — aynı ffmpeg çağrısı orijinal klibin sesini de
+map'leyip .ass altyazısını yakar. Hiç yüz bulunamazsa sabit merkez crop'a
+düşer (hata değil, sadece daha az "akıllı" bir sonuç).
+"""
+
+import asyncio
+import json
+import os
+import subprocess
+import tempfile
+from pathlib import Path
+
+import cv2
+import numpy as np
+from bullmq import Worker
+from mediapipe.tasks.python import BaseOptions
+from mediapipe.tasks.python.vision import FaceDetector, FaceDetectorOptions
+import mediapipe as mp
+
+from . import db
+from .queues import QUEUE_NAMES, REDIS_URL, SHARED_MEDIA_ROOT, VIDEO_RENDER_CONCURRENCY
+from .telegram import send_telegram_message
+
+FACE_MODEL_PATH = os.environ.get(
+    "FACE_DETECTOR_MODEL_PATH", "/app/models/blaze_face_short_range.tflite"
+)
+SAMPLE_INTERVAL_SEC = 0.5
+OUTPUT_WIDTH = 1080
+OUTPUT_HEIGHT = 1920
+
+_detector: FaceDetector | None = None
+
+
+def _get_detector() -> FaceDetector:
+    global _detector
+    if _detector is None:
+        options = FaceDetectorOptions(base_options=BaseOptions(model_asset_path=FACE_MODEL_PATH))
+        _detector = FaceDetector.create_from_options(options)
+    return _detector
+
+
+def _detect_face_centers(video_path: str) -> tuple[list[tuple[float, float]], int, int, float]:
+    """Örneklenmiş karelerde en büyük yüzün merkezini (x_frac, y_frac) döner."""
+    cap = cv2.VideoCapture(video_path)
+    fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    frame_interval = max(1, int(fps * SAMPLE_INTERVAL_SEC))
+
+    detector = _get_detector()
+    points: list[tuple[float, float]] = []  # (time_sec, x_frac)
+    frame_idx = 0
+
+    while True:
+        ok, frame = cap.read()
+        if not ok:
+            break
+        if frame_idx % frame_interval == 0:
+            rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb)
+            result = detector.detect(mp_image)
+            if result.detections:
+                best = max(
+                    result.detections,
+                    key=lambda d: d.bounding_box.width * d.bounding_box.height,
+                )
+                bbox = best.bounding_box
+                cx = (bbox.origin_x + bbox.width / 2) / width
+                points.append((frame_idx / fps, cx))
+        frame_idx += 1
+
+    cap.release()
+    return points, width, height, fps
+
+
+def _smooth_x_path(points: list[tuple[float, float]], total_frames: int, fps: float) -> np.ndarray | None:
+    """Kare başına yumuşatılmış crop merkezi (x_frac). Hiç yüz yoksa None (sabit merkez crop kullanılır)."""
+    if not points or total_frames <= 0:
+        return None
+
+    times = np.array([p[0] for p in points])
+    xs = np.array([p[1] for p in points])
+
+    frame_times = np.arange(total_frames) / fps
+    interp_x = np.interp(frame_times, times, xs)
+
+    window = min(max(1, int(fps)), len(interp_x))
+    kernel = np.ones(window) / window
+    return np.convolve(interp_x, kernel, mode="same")
+
+
+def _seconds_to_ass_time(t: float) -> str:
+    t = max(0.0, t)
+    h = int(t // 3600)
+    m = int((t % 3600) // 60)
+    s = t % 60
+    return f"{h:d}:{m:02d}:{s:05.2f}"
+
+
+def _build_ass_subtitles(transcript: dict, path: str, words_per_caption: int = 3) -> None:
+    words = transcript.get("words") or []
+    header = (
+        "[Script Info]\n"
+        "ScriptType: v4.00+\n"
+        f"PlayResX: {OUTPUT_WIDTH}\n"
+        f"PlayResY: {OUTPUT_HEIGHT}\n\n"
+        "[V4+ Styles]\n"
+        "Format: Name, Fontname, Fontsize, PrimaryColour, OutlineColour, BackColour, Bold, "
+        "BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV\n"
+        "Style: Default,Arial,64,&H00FFFFFF,&H00000000,&H80000000,1,1,3,2,2,60,60,300\n\n"
+        "[Events]\n"
+        "Format: Layer, Start, End, Style, Text\n"
+    )
+
+    lines = [header]
+    for i in range(0, len(words), words_per_caption):
+        chunk = words[i : i + words_per_caption]
+        if not chunk:
+            continue
+        start = chunk[0].get("start", 0.0)
+        end = chunk[-1].get("end", start)
+        text = " ".join(str(w.get("word", "")).strip() for w in chunk).strip()
+        if not text:
+            continue
+        lines.append(f"Dialogue: 0,{_seconds_to_ass_time(start)},{_seconds_to_ass_time(end)},Default,{text}\n")
+
+    with open(path, "w", encoding="utf-8") as f:
+        f.writelines(lines)
+
+
+def _render_final(
+    clip_path: str,
+    smoothed_x: np.ndarray | None,
+    width: int,
+    height: int,
+    fps: float,
+    ass_path: str,
+    out_path: str,
+) -> None:
+    crop_w = min(width, max(2, int(height * OUTPUT_WIDTH / OUTPUT_HEIGHT) // 2 * 2))
+
+    ffmpeg = subprocess.Popen(
+        [
+            "ffmpeg", "-y",
+            "-f", "rawvideo", "-pix_fmt", "bgr24",
+            "-s", f"{OUTPUT_WIDTH}x{OUTPUT_HEIGHT}",
+            "-r", str(fps),
+            "-i", "-",
+            "-i", clip_path,
+            "-vf", f"ass={ass_path}",
+            "-map", "0:v:0", "-map", "1:a:0?",
+            "-c:v", "libx264", "-preset", "veryfast", "-pix_fmt", "yuv420p",
+            "-c:a", "aac",
+            "-shortest",
+            out_path,
+        ],
+        stdin=subprocess.PIPE,
+        stderr=subprocess.DEVNULL,
+    )
+
+    cap = cv2.VideoCapture(clip_path)
+    frame_idx = 0
+    try:
+        while True:
+            ok, frame = cap.read()
+            if not ok:
+                break
+
+            cx_frac = 0.5
+            if smoothed_x is not None and frame_idx < len(smoothed_x):
+                cx_frac = float(smoothed_x[frame_idx])
+
+            cx_px = int(cx_frac * width)
+            x0 = max(0, min(width - crop_w, cx_px - crop_w // 2))
+            cropped = frame[0:height, x0 : x0 + crop_w]
+            resized = cv2.resize(cropped, (OUTPUT_WIDTH, OUTPUT_HEIGHT))
+
+            assert ffmpeg.stdin is not None
+            ffmpeg.stdin.write(resized.tobytes())
+            frame_idx += 1
+    finally:
+        cap.release()
+        if ffmpeg.stdin:
+            ffmpeg.stdin.close()
+        ffmpeg.wait()
+
+
+def _render_sync(file_path: str, start_sec: int, end_sec: int, transcript: dict, out_path: str) -> None:
+    with tempfile.TemporaryDirectory() as tmp_dir:
+        clip_path = os.path.join(tmp_dir, "clip.mp4")
+        subprocess.run(
+            [
+                "ffmpeg", "-y",
+                "-ss", str(start_sec),
+                "-to", str(end_sec),
+                "-i", file_path,
+                "-c", "copy",
+                clip_path,
+            ],
+            check=True,
+            capture_output=True,
+        )
+
+        points, width, height, fps = _detect_face_centers(clip_path)
+        total_frames = int(cv2.VideoCapture(clip_path).get(cv2.CAP_PROP_FRAME_COUNT))
+        smoothed = _smooth_x_path(points, total_frames, fps)
+
+        ass_path = os.path.join(tmp_dir, "subs.ass")
+        _build_ass_subtitles(transcript, ass_path)
+
+        _render_final(clip_path, smoothed, width, height, fps, ass_path, out_path)
+
+
+async def process_video_render(job, job_token=None):
+    candidate_id = job.data["candidateSegmentId"]
+    candidate = await db.get_candidate_for_render(candidate_id)
+    if candidate is None:
+        print(f"[video-render] candidate {candidate_id} not found, skipping")
+        return
+
+    short_id = await db.upsert_short_video(candidate_id, status="RENDERING")
+
+    out_dir = Path(SHARED_MEDIA_ROOT) / "shorts"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    out_path = str(out_dir / f"{candidate_id}.mp4")
+
+    try:
+        transcript = candidate["transcript_json"] or {}
+        if isinstance(transcript, str):
+            transcript = json.loads(transcript)
+
+        await asyncio.to_thread(
+            _render_sync,
+            candidate["file_path"],
+            candidate["start_sec"],
+            candidate["end_sec"],
+            transcript,
+            out_path,
+        )
+
+        await db.update_short_video(short_id, status="READY", file_path=out_path)
+        await send_telegram_message(f"🎬 *{candidate['channel_name']}* için 9:16 kısa video hazır!")
+        print(f"[video-render] short {short_id} ready: {out_path}")
+    except Exception as exc:
+        print(f"[video-render] failed for candidate {candidate_id}: {exc}")
+        await db.update_short_video(short_id, status="FAILED", error_message=str(exc)[:500])
+
+
+def start_video_render_worker() -> Worker:
+    return Worker(
+        QUEUE_NAMES["VIDEO_RENDER"],
+        process_video_render,
+        {"connection": REDIS_URL, "concurrency": VIDEO_RENDER_CONCURRENCY},
+    )
diff --git a/docker-compose.yml b/docker-compose.yml
index bbc01a4..798e415 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -65,6 +65,7 @@ services:
       TELEGRAM_BOT_TOKEN: ${TELEGRAM_BOT_TOKEN:-}
       TELEGRAM_CHAT_ID: ${TELEGRAM_CHAT_ID:-}
       WORKER_CONCURRENCY: ${WORKER_CONCURRENCY:-5}
+      VIDEO_RENDER_CONCURRENCY: ${VIDEO_RENDER_CONCURRENCY:-2}
     volumes:
       - shared-media:/shared-media
     depends_on:
diff --git a/packages/db/prisma/migrations/20260901010000_short_videos/migration.sql b/packages/db/prisma/migrations/20260901010000_short_videos/migration.sql
new file mode 100644
index 0000000..7063f73
--- /dev/null
+++ b/packages/db/prisma/migrations/20260901010000_short_videos/migration.sql
@@ -0,0 +1,21 @@
+-- CreateEnum
+CREATE TYPE "ShortVideoStatus" AS ENUM ('PENDING', 'RENDERING', 'READY', 'FAILED');
+
+-- CreateTable
+CREATE TABLE "short_videos" (
+    "id" TEXT NOT NULL,
+    "candidate_id" TEXT NOT NULL,
+    "file_path" TEXT,
+    "status" "ShortVideoStatus" NOT NULL DEFAULT 'PENDING',
+    "error_message" TEXT,
+    "created_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    "updated_at" TIMESTAMP(3) NOT NULL,
+
+    CONSTRAINT "short_videos_pkey" PRIMARY KEY ("id")
+);
+
+-- CreateIndex
+CREATE UNIQUE INDEX "short_videos_candidate_id_key" ON "short_videos"("candidate_id");
+
+-- AddForeignKey
+ALTER TABLE "short_videos" ADD CONSTRAINT "short_videos_candidate_id_fkey" FOREIGN KEY ("candidate_id") REFERENCES "candidate_segments"("id") ON DELETE RESTRICT ON UPDATE CASCADE;
diff --git a/packages/db/prisma/schema.prisma b/packages/db/prisma/schema.prisma
index e5de72c..cf9a7c3 100644
--- a/packages/db/prisma/schema.prisma
+++ b/packages/db/prisma/schema.prisma
@@ -19,6 +19,13 @@ enum CandidateSegmentStatus {
   FAILED
 }
 
+enum ShortVideoStatus {
+  PENDING
+  RENDERING
+  READY
+  FAILED
+}
+
 model Channel {
   id             String    @id @default(cuid())
   name           String
@@ -74,9 +81,24 @@ model CandidateSegment {
   status           CandidateSegmentStatus  @default(PENDING_STT)
   createdAt        DateTime                @default(now()) @map("created_at")
 
+  short ShortVideo?
+
   @@map("candidate_segments")
 }
 
+model ShortVideo {
+  id           String           @id @default(cuid())
+  candidateId  String           @unique @map("candidate_id")
+  candidate    CandidateSegment @relation(fields: [candidateId], references: [id])
+  filePath     String?          @map("file_path")
+  status       ShortVideoStatus @default(PENDING)
+  errorMessage String?          @map("error_message")
+  createdAt    DateTime         @default(now()) @map("created_at")
+  updatedAt    DateTime         @updatedAt @map("updated_at")
+
+  @@map("short_videos")
+}
+
 model AppSetting {
   key       String   @id
   value     String