diff --git a/apps/web/Dockerfile b/apps/web/Dockerfile
index 380496bab5..d9b09f2b48 100644
--- a/apps/web/Dockerfile
+++ b/apps/web/Dockerfile
@@ -17,6 +17,14 @@ WORKDIR /app
 
 ARG FREESOUND_CLIENT_ID
 ARG FREESOUND_API_KEY
+# NEXT_PUBLIC_* are inlined at build time — pass VPS URLs as build args.
+ARG NEXT_PUBLIC_SITE_URL=http://localhost:3000
+ARG NEXT_PUBLIC_MARBLE_API_URL=https://api.marblecms.com
+ARG NEXT_PUBLIC_AI_BACKEND_URL=http://localhost:8420
+ARG NEXT_PUBLIC_OLLAMA_URL=http://localhost:11434
+ARG NEXT_PUBLIC_WHISPER_SERVICE_URL=http://localhost:8421
+ARG NEXT_PUBLIC_TTS_SERVICE_URL=http://localhost:8422
+ARG NEXT_PUBLIC_IMAGE_SERVICE_URL=http://localhost:8423
 
 COPY package.json package.json
 COPY bun.lock bun.lock
@@ -43,9 +51,13 @@ ENV DATABASE_URL="postgresql://opencut:opencut@localhost:5432/opencut"
 ENV BETTER_AUTH_SECRET="build-time-secret"
 ENV UPSTASH_REDIS_REST_URL="http://localhost:8079"
 ENV UPSTASH_REDIS_REST_TOKEN="example_token"
-ENV NEXT_PUBLIC_SITE_URL="http://localhost:3000"
-ENV NEXT_PUBLIC_MARBLE_API_URL="https://api.marblecms.com"
-ENV NEXT_PUBLIC_AI_BACKEND_URL="http://localhost:8420"
+ENV NEXT_PUBLIC_SITE_URL=$NEXT_PUBLIC_SITE_URL
+ENV NEXT_PUBLIC_MARBLE_API_URL=$NEXT_PUBLIC_MARBLE_API_URL
+ENV NEXT_PUBLIC_AI_BACKEND_URL=$NEXT_PUBLIC_AI_BACKEND_URL
+ENV NEXT_PUBLIC_OLLAMA_URL=$NEXT_PUBLIC_OLLAMA_URL
+ENV NEXT_PUBLIC_WHISPER_SERVICE_URL=$NEXT_PUBLIC_WHISPER_SERVICE_URL
+ENV NEXT_PUBLIC_TTS_SERVICE_URL=$NEXT_PUBLIC_TTS_SERVICE_URL
+ENV NEXT_PUBLIC_IMAGE_SERVICE_URL=$NEXT_PUBLIC_IMAGE_SERVICE_URL
 
 ENV FREESOUND_CLIENT_ID=$FREESOUND_CLIENT_ID
 ENV FREESOUND_API_KEY=$FREESOUND_API_KEY
diff --git a/apps/web/src/app/layout.tsx b/apps/web/src/app/layout.tsx
index 981cb201c2..a0c8c3bf28 100644
--- a/apps/web/src/app/layout.tsx
+++ b/apps/web/src/app/layout.tsx
@@ -48,18 +48,20 @@ export default function RootLayout({
 				>
 					<TooltipProvider>
 						<Toaster />
-						<Script
-							src="https://cdn.databuddy.cc/databuddy.js"
-							strategy="afterInteractive"
-							async
-							data-client-id="UP-Wcoy5arxFeK7oyjMMZ"
-							data-disabled={webEnv.NODE_ENV === "development"}
-							data-track-attributes={false}
-							data-track-errors={true}
-							data-track-outgoing-links={false}
-							data-track-web-vitals={false}
-							data-track-sessions={false}
-						/>
+						{process.env.NEXT_PUBLIC_DATABUDDY_CLIENT_ID && (
+							<Script
+								src="https://cdn.databuddy.cc/databuddy.js"
+								strategy="afterInteractive"
+								async
+								data-client-id={process.env.NEXT_PUBLIC_DATABUDDY_CLIENT_ID}
+								data-disabled={webEnv.NODE_ENV === "development"}
+								data-track-attributes={false}
+								data-track-errors={true}
+								data-track-outgoing-links={false}
+								data-track-web-vitals={false}
+								data-track-sessions={false}
+							/>
+						)}
 						{children}
 					</TooltipProvider>
 				</ThemeProvider>
diff --git a/apps/web/src/components/editor/ai/ai-panel-wrapper.tsx b/apps/web/src/components/editor/ai/ai-panel-wrapper.tsx
index f4ff83f7c8..be91396548 100644
--- a/apps/web/src/components/editor/ai/ai-panel-wrapper.tsx
+++ b/apps/web/src/components/editor/ai/ai-panel-wrapper.tsx
@@ -12,6 +12,7 @@ import { cn } from "@/utils/ui";
 import { Button } from "@/components/ui/button";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { SparklesIcon, Cancel01Icon } from "@hugeicons/core-free-icons";
+import { generateUUID } from "@/utils/id";
 
 export function AIPanelWrapper() {
 	const {
@@ -39,7 +40,7 @@ export function AIPanelWrapper() {
 		async (message: string) => {
 			if (!isConnected) {
 				const errorMessage: AIMessage = {
-					id: crypto.randomUUID(),
+					id: generateUUID(),
 					role: "assistant",
 					content:
 						"Cannot process commands — the AI backend is not connected. Open the setup guide from the AI status indicator in the header to get started.",
@@ -50,7 +51,7 @@ export function AIPanelWrapper() {
 			}
 
 			const userMessage: AIMessage = {
-				id: crypto.randomUUID(),
+				id: generateUUID(),
 				role: "user",
 				content: message,
 				timestamp: new Date(),
@@ -61,12 +62,12 @@ export function AIPanelWrapper() {
 				const result = await executeCommand(message, null);
 
 				const assistantMessage: AIMessage = {
-					id: crypto.randomUUID(),
+					id: generateUUID(),
 					role: "assistant",
 					content: result.explanation,
 					timestamp: new Date(),
 					actions: result.actions.map((action) => ({
-						id: crypto.randomUUID(),
+						id: generateUUID(),
 						type: action.type,
 						label: action.type.replace(/_/g, " ").toLowerCase(),
 						description: action.description,
@@ -84,7 +85,7 @@ export function AIPanelWrapper() {
 					errorDetail.includes("Failed to fetch");
 
 				const errorMessage: AIMessage = {
-					id: crypto.randomUUID(),
+					id: generateUUID(),
 					role: "assistant",
 					content: isConnectionIssue
 						? `Cannot reach the AI backend: ${errorDetail}\n\nMake sure the AI backend is running at the configured URL. Open the setup guide for instructions.`
diff --git a/apps/web/src/components/editor/panels/assets/views/ai-studio.tsx b/apps/web/src/components/editor/panels/assets/views/ai-studio.tsx
index a24262e891..e64ff08cae 100644
--- a/apps/web/src/components/editor/panels/assets/views/ai-studio.tsx
+++ b/apps/web/src/components/editor/panels/assets/views/ai-studio.tsx
@@ -22,6 +22,7 @@ import { useAIStatus } from "@/hooks/use-ai-status";
 import { useAIStore } from "@/stores/ai-store";
 import { useTranscriptStore } from "@/stores/transcript-store";
 import { toast } from "sonner";
+import { generateUUID } from "@/utils/id";
 import { TemplatePanel } from "@/components/editor/ai/template-panel";
 import { BRollSuggestionsPanel } from "@/components/editor/ai/broll-suggestions-panel";
 import { YouTubeReelsPanel } from "@/components/editor/youtube/youtube-reels-panel";
@@ -364,14 +365,14 @@ export function AIStudioView() {
 		}
 
 		addMessage({
-			id: crypto.randomUUID(),
+			id: generateUUID(),
 			role: "user",
 			content: trimmed,
 		});
 		setInputValue("");
 		setIsThinking(true);
 
-		const assistantId = crypto.randomUUID();
+		const assistantId = generateUUID();
 		let messageAdded = false;
 
 		try {
diff --git a/apps/web/src/components/editor/panels/timeline/audio-effects-panel.tsx b/apps/web/src/components/editor/panels/timeline/audio-effects-panel.tsx
index 8dd84fa6e3..71181c1a67 100644
--- a/apps/web/src/components/editor/panels/timeline/audio-effects-panel.tsx
+++ b/apps/web/src/components/editor/panels/timeline/audio-effects-panel.tsx
@@ -9,14 +9,11 @@ import { Slider } from "@/components/ui/slider";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Add01Icon, Cancel01Icon } from "@hugeicons/core-free-icons";
 import { useEditor } from "@/hooks/use-editor";
-import { getAllAudioEffectDefinitions, type AudioEffectType } from "@/lib/audio/audio-effects";
-
-interface TrackAudioEffect {
-	id: string;
-	type: AudioEffectType;
-	params: Record<string, number>;
-	enabled: boolean;
-}
+import {
+	getAllAudioEffectDefinitions,
+	type AudioEffectType,
+	type TrackAudioEffect,
+} from "@/lib/audio/audio-effects";
 
 interface AudioEffectsChainProps {
 	trackId: string;
@@ -29,7 +26,10 @@ export function AudioEffectsChainPanel({ trackId, trackType, className }: AudioE
 	const track = editor.timeline.getTrackById({ trackId });
 	if (!track || (trackType !== "audio" && trackType !== "video")) return null;
 
-	const effects: TrackAudioEffect[] = (track as any).audioEffects ?? [];
+	const effects: TrackAudioEffect[] =
+		(track.type === "audio" || track.type === "video"
+			? track.audioEffects
+			: undefined) ?? [];
 
 	const handleAdd = useCallback(
 		(type: AudioEffectType) => {
@@ -45,7 +45,7 @@ export function AudioEffectsChainPanel({ trackId, trackType, className }: AudioE
 			];
 			editor.timeline.updateTrack({
 				trackId,
-				updates: { audioEffects: updatedEffects } as any,
+				updates: { audioEffects: updatedEffects },
 			});
 		},
 		[effects, trackId, editor],
@@ -56,7 +56,7 @@ export function AudioEffectsChainPanel({ trackId, trackType, className }: AudioE
 			const updatedEffects = effects.filter((e) => e.id !== effectId);
 			editor.timeline.updateTrack({
 				trackId,
-				updates: { audioEffects: updatedEffects } as any,
+				updates: { audioEffects: updatedEffects },
 			});
 		},
 		[effects, trackId, editor],
@@ -69,7 +69,7 @@ export function AudioEffectsChainPanel({ trackId, trackType, className }: AudioE
 			);
 			editor.timeline.updateTrack({
 				trackId,
-				updates: { audioEffects: updatedEffects } as any,
+				updates: { audioEffects: updatedEffects },
 			});
 		},
 		[effects, trackId, editor],
@@ -82,7 +82,7 @@ export function AudioEffectsChainPanel({ trackId, trackType, className }: AudioE
 			);
 			editor.timeline.updateTrack({
 				trackId,
-				updates: { audioEffects: updatedEffects } as any,
+				updates: { audioEffects: updatedEffects },
 			});
 		},
 		[effects, trackId, editor],
diff --git a/apps/web/src/components/editor/panels/timeline/audio-mixer-panel.tsx b/apps/web/src/components/editor/panels/timeline/audio-mixer-panel.tsx
index e4fac4718a..0df08e4d3e 100644
--- a/apps/web/src/components/editor/panels/timeline/audio-mixer-panel.tsx
+++ b/apps/web/src/components/editor/panels/timeline/audio-mixer-panel.tsx
@@ -4,6 +4,12 @@ import { useCallback, useEffect, useRef, useState } from "react";
 import { useEditor } from "@/hooks/use-editor";
 import { cn } from "@/utils/ui";
 import type { TimelineTrack, VideoTrack, AudioTrack } from "@/types/timeline";
+import {
+	Popover,
+	PopoverContent,
+	PopoverTrigger,
+} from "@/components/ui/popover";
+import { AudioEffectsChainPanel } from "./audio-effects-panel";
 
 type AudioTrackish = VideoTrack | AudioTrack;
 
@@ -113,6 +119,7 @@ function TrackMixerStrip({
 	const pan = "pan" in track ? track.pan ?? 0 : 0;
 	const isMuted = track.muted;
 	const isSolo = track.solo ?? false;
+	const effectCount = track.audioEffects?.length ?? 0;
 
 	const handleVolumeChange = useCallback(
 		(e: React.ChangeEvent<HTMLInputElement>) => {
@@ -229,6 +236,25 @@ function TrackMixerStrip({
 					</span>
 				</div>
 			)}
+
+			<Popover>
+				<PopoverTrigger asChild>
+					<button
+						type="button"
+						className={cn(
+							"w-full rounded text-[8px] font-medium py-0.5 border",
+							effectCount > 0
+								? "border-primary/50 text-primary bg-primary/10"
+								: "border-border text-muted-foreground",
+						)}
+					>
+						FX{effectCount > 0 ? ` (${effectCount})` : ""}
+					</button>
+				</PopoverTrigger>
+				<PopoverContent side="top" align="center" className="w-56 p-2">
+					<AudioEffectsChainPanel trackId={track.id} trackType={track.type} />
+				</PopoverContent>
+			</Popover>
 		</div>
 	);
 }
diff --git a/apps/web/src/core/managers/audio-manager.ts b/apps/web/src/core/managers/audio-manager.ts
index 5b24dd28d7..3804d37aea 100644
--- a/apps/web/src/core/managers/audio-manager.ts
+++ b/apps/web/src/core/managers/audio-manager.ts
@@ -1,6 +1,10 @@
 import type { EditorCore } from "@/core";
 import type { AudioClipSource } from "@/lib/media/audio";
 import { createAudioContext, collectAudioClips } from "@/lib/media/audio";
+import {
+	getAudioEffectDefinition,
+	type TrackAudioEffect,
+} from "@/lib/audio/audio-effects";
 import {
 	ALL_FORMATS,
 	AudioBufferSink,
@@ -154,6 +158,9 @@ export class AudioManager {
 			(track && ("volume" in track ? track.volume : undefined)) ?? 1;
 		const trackPan =
 			(track && ("pan" in track ? track.pan : undefined)) ?? 0;
+		const audioEffects =
+			(track && "audioEffects" in track ? track.audioEffects : undefined) ??
+			[];
 
 		const gain = ctx.createGain();
 		gain.gain.value = trackVolume;
@@ -165,7 +172,12 @@ export class AudioManager {
 		analyser.fftSize = 256;
 		analyser.smoothingTimeConstant = 0.8;
 
-		gain.connect(panner);
+		this.connectAudioEffectsChain({
+			ctx,
+			input: gain,
+			output: panner,
+			effects: audioEffects,
+		});
 		panner.connect(analyser);
 		analyser.connect(this.masterGain);
 
@@ -174,6 +186,39 @@ export class AudioManager {
 		return nodes;
 	}
 
+	// wires enabled track audio effects (EQ, compressor, reverb, etc.) in
+	// series between the track gain and panner nodes
+	private connectAudioEffectsChain({
+		ctx,
+		input,
+		output,
+		effects,
+	}: {
+		ctx: AudioContext;
+		input: AudioNode;
+		output: AudioNode;
+		effects: TrackAudioEffect[];
+	}): void {
+		const enabledEffects = effects.filter((effect) => effect.enabled);
+		if (enabledEffects.length === 0) {
+			input.connect(output);
+			return;
+		}
+
+		let previousOutput: AudioNode = input;
+		for (const effect of enabledEffects) {
+			const definition = getAudioEffectDefinition(effect.type);
+			if (!definition) continue;
+
+			const nodes = definition.createNodes(ctx, effect.params);
+			if (nodes.length === 0) continue;
+
+			previousOutput.connect(nodes[0]);
+			previousOutput = nodes[nodes.length - 1];
+		}
+		previousOutput.connect(output);
+	}
+
 	private rebuildTrackNodes(): void {
 		for (const [, nodes] of this.trackNodes) {
 			try {
diff --git a/apps/web/src/core/managers/playback-manager.ts b/apps/web/src/core/managers/playback-manager.ts
index e86c883659..5ef813dafa 100644
--- a/apps/web/src/core/managers/playback-manager.ts
+++ b/apps/web/src/core/managers/playback-manager.ts
@@ -27,6 +27,12 @@ export class PlaybackManager {
 			this.seek({ time: 0 });
 		}
 
+		// Normal play is always 1x. Shuttle sets its own direction/speed via
+		// shuttleForward/shuttleReverse; leaving speed at 0 made time never
+		// advance and, at t=0, every rAF hit newTime <= 0 and stormed
+		// playback-seek events (audio restart / play-pause bootloop).
+		this.shuttleSpeed = 0;
+		this.shuttleDirection = null;
 		this.isPlaying = true;
 		this.startTimer();
 		this.notify();
@@ -34,6 +40,8 @@ export class PlaybackManager {
 
 	pause(): void {
 		this.isPlaying = false;
+		this.shuttleSpeed = 0;
+		this.shuttleDirection = null;
 		this.stopTimer();
 		this.notify();
 	}
@@ -192,7 +200,14 @@ export class PlaybackManager {
 		const delta = (now - this.lastUpdate) / 1000;
 		this.lastUpdate = now;
 
-		const speed = this.shuttleDirection === "reverse" ? -this.shuttleSpeed : this.shuttleSpeed;
+		// Shuttle inactive → 1x (normal play). Active shuttle uses its speed.
+		let speed = 1;
+		if (this.shuttleDirection === "forward") {
+			speed = this.shuttleSpeed;
+		} else if (this.shuttleDirection === "reverse") {
+			speed = -this.shuttleSpeed;
+		}
+
 		const newTime = this.currentTime + delta * speed;
 		const duration = this.editor.timeline.getTotalDuration();
 
@@ -206,18 +221,26 @@ export class PlaybackManager {
 					detail: { time: duration },
 				}),
 			);
-		} else if (newTime <= 0) {
+			return;
+		}
+
+		if (newTime <= 0) {
+			const wasAtStart = this.currentTime <= 0;
 			this.currentTime = 0;
 			this.notify();
 
-			window.dispatchEvent(
-				new CustomEvent("playback-seek", {
-					detail: { time: 0 },
-				}),
-			);
+			// Avoid seek storms when stuck at 0 (e.g. speed was 0 before the fix)
+			if (!wasAtStart) {
+				window.dispatchEvent(
+					new CustomEvent("playback-seek", {
+						detail: { time: 0 },
+					}),
+				);
+			}
 
 			if (this.shuttleDirection === "reverse" && this.shuttleSpeed > 0) {
 				this.shuttleStop();
+				return;
 			}
 		} else {
 			this.currentTime = newTime;
diff --git a/apps/web/src/core/managers/timeline-manager.ts b/apps/web/src/core/managers/timeline-manager.ts
index 5e6d237f5c..d8956504a1 100644
--- a/apps/web/src/core/managers/timeline-manager.ts
+++ b/apps/web/src/core/managers/timeline-manager.ts
@@ -1,5 +1,6 @@
 import type { EditorCore } from "@/core";
 import type { EffectParamValues } from "@/types/effects";
+import type { TrackAudioEffect } from "@/lib/audio/audio-effects";
 import type {
 	TrackType,
 	TimelineTrack,
@@ -619,7 +620,16 @@ export class TimelineManager {
 		updates,
 	}: {
 		trackId: string;
-		updates: Partial<{ muted: boolean; hidden: boolean; volume: number; pan: number; solo: boolean; color: string; locked: boolean }>;
+		updates: Partial<{
+			muted: boolean;
+			hidden: boolean;
+			volume: number;
+			pan: number;
+			solo: boolean;
+			color: string;
+			locked: boolean;
+			audioEffects: TrackAudioEffect[];
+		}>;
 	}): void {
 		const tracks = this.getTracks();
 		const updatedTracks = tracks.map((track) =>
diff --git a/apps/web/src/hooks/use-service-health.ts b/apps/web/src/hooks/use-service-health.ts
index 3d18ac6eb1..9056eb92f8 100644
--- a/apps/web/src/hooks/use-service-health.ts
+++ b/apps/web/src/hooks/use-service-health.ts
@@ -29,7 +29,7 @@ export interface AllServicesHealth {
 
 export const SERVICE_URLS = {
 	backend: process.env.NEXT_PUBLIC_AI_BACKEND_URL || "http://localhost:8420",
-	ollama: "http://localhost:11434",
+	ollama: process.env.NEXT_PUBLIC_OLLAMA_URL || "http://localhost:11434",
 	whisper: process.env.NEXT_PUBLIC_WHISPER_SERVICE_URL || "http://localhost:8421",
 	tts: process.env.NEXT_PUBLIC_TTS_SERVICE_URL || "http://localhost:8422",
 	image: process.env.NEXT_PUBLIC_IMAGE_SERVICE_URL || "http://localhost:8423",
diff --git a/apps/web/src/lib/audio/audio-effects.ts b/apps/web/src/lib/audio/audio-effects.ts
index 1cce8f931a..bbeb69b88c 100644
--- a/apps/web/src/lib/audio/audio-effects.ts
+++ b/apps/web/src/lib/audio/audio-effects.ts
@@ -25,6 +25,14 @@ export interface AudioEffectDefinition {
 	createNodes: (ctx: AudioContext, params: Record<string, number>) => AudioNode[];
 }
 
+/** A configured audio effect instance stored on a track. */
+export interface TrackAudioEffect {
+	id: string;
+	type: AudioEffectType;
+	params: Record<string, number>;
+	enabled: boolean;
+}
+
 export const AUDIO_EFFECT_DEFINITIONS: Record<AudioEffectType, AudioEffectDefinition> = {
 	eq: {
 		type: "eq",
@@ -40,7 +48,7 @@ export const AUDIO_EFFECT_DEFINITIONS: Record<AudioEffectType, AudioEffectDefini
 		createNodes: (ctx, params) => {
 			const low = ctx.createBiquadFilter();
 			low.type = "lowshelf";
-			low.frequency.value = params.lowGain !== undefined ? (params.lowFreq as number) : 320;
+			low.frequency.value = params.lowFreq ?? 320;
 			low.gain.value = params.lowGain ?? 0;
 
 			const mid = ctx.createBiquadFilter();
@@ -113,6 +121,11 @@ export const AUDIO_EFFECT_DEFINITIONS: Record<AudioEffectType, AudioEffectDefini
 			{ key: "wetLevel", label: "Wet Level", type: "number", default: 30, min: 0, max: 100, step: 1, unit: "%" },
 		],
 		createNodes: (ctx, params) => {
+			// unity fan-out node so both the wet (convolver) and dry paths
+			// receive the same input signal — a single node can't be
+			// connected to twice as a source, so callers must connect to
+			// this node's input and read output from the merger
+			const input = ctx.createGain();
 			const convolver = ctx.createConvolver();
 			const wet = ctx.createGain();
 			const dry = ctx.createGain();
@@ -136,13 +149,15 @@ export const AUDIO_EFFECT_DEFINITIONS: Record<AudioEffectType, AudioEffectDefini
 
 			convolver.buffer = impulse;
 			wet.gain.value = wetLevel;
-			dry.gain.value = 1;
+			dry.gain.value = 1 - wetLevel;
 
+			input.connect(convolver);
+			input.connect(dry);
 			convolver.connect(wet);
 			wet.connect(merger);
 			dry.connect(merger);
 
-			return [convolver, merger];
+			return [input, merger];
 		},
 	},
 	"de-esser": {
diff --git a/apps/web/src/lib/effects/definitions/sharpen.frag.glsl b/apps/web/src/lib/effects/definitions/sharpen.frag.glsl
index cd6ad877d2..43ae962891 100644
--- a/apps/web/src/lib/effects/definitions/sharpen.frag.glsl
+++ b/apps/web/src/lib/effects/definitions/sharpen.frag.glsl
@@ -2,11 +2,21 @@ precision mediump float;
 
 varying vec2 v_texCoord;
 uniform sampler2D u_texture;
+uniform vec2 u_resolution;
 uniform float u_amount;
 
 void main() {
+  vec2 texel = 1.0 / u_resolution;
   vec4 color = texture2D(u_texture, v_texCoord);
-  vec3 rgb = color.rgb;
-  rgb = rgb * (2.0 * u_amount + 1.0) / (rgb * 2.0 * u_amount + 2.0 * u_amount - rgb * 2.0 * u_amount + 1.0);
-  gl_FragColor = vec4(clamp(rgb, 0.0, 1.0), color.a);
+
+  // classic unsharp mask: boost the center pixel by its difference from
+  // the average of its 4-neighborhood, scaled by amount
+  vec4 neighborSum =
+      texture2D(u_texture, v_texCoord + vec2(-texel.x, 0.0)) +
+      texture2D(u_texture, v_texCoord + vec2(texel.x, 0.0)) +
+      texture2D(u_texture, v_texCoord + vec2(0.0, -texel.y)) +
+      texture2D(u_texture, v_texCoord + vec2(0.0, texel.y));
+
+  vec3 sharpened = color.rgb * (1.0 + 4.0 * u_amount) - neighborSum.rgb * u_amount;
+  gl_FragColor = vec4(clamp(sharpened, 0.0, 1.0), color.a);
 }
diff --git a/apps/web/src/services/renderer/nodes/transition-node.ts b/apps/web/src/services/renderer/nodes/transition-node.ts
index 599c8e0bb2..d5e249c53b 100644
--- a/apps/web/src/services/renderer/nodes/transition-node.ts
+++ b/apps/web/src/services/renderer/nodes/transition-node.ts
@@ -1,15 +1,12 @@
 import type { CanvasRenderer } from "../canvas-renderer";
 import { createOffscreenCanvas } from "../canvas-utils";
 import { BaseNode } from "./base-node";
-import type { Transform, TransitionData } from "@/types/timeline";
+import { VideoNode } from "./video-node";
+import { ImageNode } from "./image-node";
+import type { Transform } from "@/types/timeline";
 import type { Effect } from "@/types/effects";
 import type { BlendMode } from "@/types/rendering";
 import type { ElementAnimations } from "@/types/animation";
-import {
-	resolveOpacityAtTime,
-	resolveTransformAtTime,
-	getElementLocalTime,
-} from "@/lib/animation";
 import { renderTransition } from "../transition-renderer";
 import { getTransition } from "@/lib/transitions";
 
@@ -24,6 +21,9 @@ export interface TransitionSourceParams {
 	opacity: number;
 	blendMode?: BlendMode;
 	effects?: Effect[];
+	/** Media backing this side of the transition, if any (video/image only). */
+	mediaId?: string;
+	mediaType?: "video" | "image";
 }
 
 export interface TransitionNodeParams {
@@ -33,8 +33,6 @@ export interface TransitionNodeParams {
 	sourceA: TransitionSourceParams;
 	sourceB: TransitionSourceParams;
 	mediaMap: Map<string, { url: string; file?: File }>;
-	mediaIdA?: string;
-	mediaIdB?: string;
 }
 
 export class TransitionNode extends BaseNode<TransitionNodeParams> {
@@ -54,12 +52,12 @@ export class TransitionNode extends BaseNode<TransitionNodeParams> {
 
 		const progress = (time - transitionStart) / transitionDuration;
 
-		const canvasA = this.renderSourceFrame({
+		const canvasA = await this.renderSourceFrame({
 			renderer,
 			sourceParams: this.params.sourceA,
 			time,
 		});
-		const canvasB = this.renderSourceFrame({
+		const canvasB = await this.renderSourceFrame({
 			renderer,
 			sourceParams: this.params.sourceB,
 			time,
@@ -80,11 +78,50 @@ export class TransitionNode extends BaseNode<TransitionNodeParams> {
 
 		renderer.context.save();
 		renderer.context.globalCompositeOperation = "source-over";
+		renderer.context.globalAlpha = 1;
 		renderer.context.drawImage(result as CanvasImageSource, 0, 0);
 		renderer.context.restore();
 	}
 
-	private renderSourceFrame({
+	private buildContentNode({
+		sourceParams,
+	}: {
+		sourceParams: TransitionSourceParams;
+	}): BaseNode | null {
+		if (!sourceParams.mediaId || !sourceParams.mediaType) return null;
+
+		const media = this.params.mediaMap.get(sourceParams.mediaId);
+		if (!media?.url) return null;
+
+		const shared = {
+			duration: sourceParams.duration,
+			timeOffset: sourceParams.timeOffset,
+			trimStart: sourceParams.trimStart,
+			trimEnd: sourceParams.trimEnd,
+			playbackRate: sourceParams.playbackRate,
+			transform: sourceParams.transform,
+			animations: sourceParams.animations,
+			opacity: sourceParams.opacity,
+			blendMode: sourceParams.blendMode,
+			effects: sourceParams.effects,
+		};
+
+		if (sourceParams.mediaType === "video") {
+			if (!media.file) return null;
+			return new VideoNode({
+				...shared,
+				url: media.url,
+				file: media.file,
+				mediaId: sourceParams.mediaId,
+			});
+		}
+
+		return new ImageNode({ ...shared, url: media.url });
+	}
+
+	// renders the source clip's frame (respecting transform/trim/effects) into
+	// its own offscreen canvas so the transition shader has real pixels to blend
+	private async renderSourceFrame({
 		renderer,
 		sourceParams,
 		time,
@@ -92,7 +129,7 @@ export class TransitionNode extends BaseNode<TransitionNodeParams> {
 		renderer: CanvasRenderer;
 		sourceParams: TransitionSourceParams;
 		time: number;
-	}): HTMLCanvasElement | OffscreenCanvas | null {
+	}): Promise<HTMLCanvasElement | OffscreenCanvas | null> {
 		const offscreen = createOffscreenCanvas({
 			width: renderer.width,
 			height: renderer.height,
@@ -103,24 +140,16 @@ export class TransitionNode extends BaseNode<TransitionNodeParams> {
 			| null;
 		if (!ctx) return null;
 
-		const animLocalTime = getElementLocalTime({
-			timelineTime: time,
-			elementStartTime: sourceParams.timeOffset,
-			elementDuration: sourceParams.duration,
-		});
-
-		const transform = resolveTransformAtTime({
-			baseTransform: sourceParams.transform,
-			animations: sourceParams.animations,
-			localTime: animLocalTime,
-		});
-		const opacity = resolveOpacityAtTime({
-			baseOpacity: sourceParams.opacity,
-			animations: sourceParams.animations,
-			localTime: animLocalTime,
-		});
+		const contentNode = this.buildContentNode({ sourceParams });
+		if (!contentNode) return offscreen;
 
-		(ctx as CanvasRenderingContext2D).globalAlpha = opacity;
+		const originalContext = renderer.context;
+		renderer.context = ctx;
+		try {
+			await contentNode.render({ renderer, time });
+		} finally {
+			renderer.context = originalContext;
+		}
 
 		return offscreen;
 	}
diff --git a/apps/web/src/services/renderer/scene-builder.ts b/apps/web/src/services/renderer/scene-builder.ts
index 44fb9254c6..18134fe577 100644
--- a/apps/web/src/services/renderer/scene-builder.ts
+++ b/apps/web/src/services/renderer/scene-builder.ts
@@ -1,4 +1,9 @@
-import type { TimelineTrack, VisualElement } from "@/types/timeline";
+import type {
+	ImageElement,
+	TimelineTrack,
+	VideoElement,
+	VisualElement,
+} from "@/types/timeline";
 import type { MediaAsset } from "@/types/assets";
 import { RootNode } from "./nodes/root-node";
 import { VideoNode } from "./nodes/video-node";
@@ -275,10 +280,7 @@ function buildTransitionNodes({
 				const asset = mediaMap.get(current.mediaId);
 				if (!asset) continue;
 
-				const nextAsset =
-					next.type === "video" || next.type === "image"
-						? mediaMap.get(next.mediaId)
-						: null;
+				const nextIsMedia = next.type === "video" || next.type === "image";
 
 				transitionNodes.push(
 					new TransitionNode({
@@ -297,6 +299,8 @@ function buildTransitionNodes({
 							opacity: current.opacity,
 							blendMode: current.blendMode,
 							effects: current.effects,
+							mediaId: current.mediaId,
+							mediaType: current.type,
 						},
 						sourceB: {
 							duration: next.duration,
@@ -310,16 +314,17 @@ function buildTransitionNodes({
 							opacity: (next as VisualElement).opacity,
 							blendMode: (next as VisualElement).blendMode,
 							effects: (next as VisualElement).effects,
+							mediaId: nextIsMedia
+								? (next as VideoElement | ImageElement).mediaId
+								: undefined,
+							mediaType: nextIsMedia
+								? (next.type as "video" | "image")
+								: undefined,
 						},
 						mediaMap: mediaMap as unknown as Map<
 							string,
 							{ url: string; file?: File }
 						>,
-						mediaIdA: current.mediaId,
-						mediaIdB:
-							next.type === "video" || next.type === "image"
-								? next.mediaId
-								: undefined,
 					}),
 				);
 			}
diff --git a/apps/web/src/services/version/ai-commit-message.ts b/apps/web/src/services/version/ai-commit-message.ts
index ecb2d164c0..17c6c3314e 100644
--- a/apps/web/src/services/version/ai-commit-message.ts
+++ b/apps/web/src/services/version/ai-commit-message.ts
@@ -7,27 +7,19 @@ const AI_BACKEND_URL =
 
 /**
  * Generate an AI-powered commit message from a diff.
- * Tries the local AI backend first, falls back to Ollama.
+ * All LLM traffic goes through the AI backend (which proxies Ollama);
+ * the browser never talks to Ollama directly.
  */
 export async function generateAICommitMessage(
 	diff: TimelineDiff,
 ): Promise<string | null> {
 	const prompt = buildPrompt(diff);
 
-	// Try AI backend
 	try {
 		const message = await tryAIBackend(prompt);
 		if (message) return message;
 	} catch {
-		// Fall through to Ollama
-	}
-
-	// Try Ollama
-	try {
-		const message = await tryOllama(prompt);
-		if (message) return message;
-	} catch {
-		// Both failed
+		// Backend unavailable — no commit message
 	}
 
 	return null;
@@ -80,48 +72,31 @@ async function tryAIBackend(prompt: string): Promise<string | null> {
 		const response = await fetch(`${AI_BACKEND_URL}/api/llm/chat/stream`, {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
-			body: JSON.stringify({
-				messages: [{ role: "user", content: prompt }],
-				stream: false,
-			}),
+			body: JSON.stringify({ message: prompt }),
 			signal: controller.signal,
 		});
 
 		if (!response.ok) return null;
 
-		const data = await response.json();
-		const text = data?.response || data?.message?.content || data?.choices?.[0]?.message?.content;
+		// Response is newline-delimited JSON: {"token": "..."} then {"done": true}
+		const raw = await response.text();
+		let text = "";
+		for (const line of raw.split("\n")) {
+			if (!line.trim()) continue;
+			try {
+				const chunk = JSON.parse(line);
+				if (typeof chunk.token === "string") text += chunk.token;
+				if (chunk.error) return null;
+			} catch {
+				// Skip malformed lines
+			}
+		}
 		return text ? cleanCommitMessage(text) : null;
 	} finally {
 		clearTimeout(timeout);
 	}
 }
 
-async function tryOllama(prompt: string): Promise<string | null> {
-	const controller = new AbortController();
-	const timeout = setTimeout(() => controller.abort(), 15000);
-
-	try {
-		const response = await fetch("http://localhost:11434/api/generate", {
-			method: "POST",
-			headers: { "Content-Type": "application/json" },
-			body: JSON.stringify({
-				model: "llama3.2:1b",
-				prompt,
-				stream: false,
-			}),
-			signal: controller.signal,
-		});
-
-		if (!response.ok) return null;
-
-		const data = await response.json();
-		return data?.response ? cleanCommitMessage(data.response) : null;
-	} finally {
-		clearTimeout(timeout);
-	}
-}
-
 function cleanCommitMessage(raw: string): string {
 	return raw
 		.replace(/^["']|["']$/g, "") // Remove surrounding quotes
diff --git a/apps/web/src/stores/ai-store.ts b/apps/web/src/stores/ai-store.ts
index 62deaf8430..91d97de8d4 100644
--- a/apps/web/src/stores/ai-store.ts
+++ b/apps/web/src/stores/ai-store.ts
@@ -1,5 +1,6 @@
 import { create } from "zustand";
 import type { AIBackendStatus, AIErrorType, AISuggestion } from "@/types/ai";
+import { generateUUID } from "@/utils/id";
 
 export interface SavedIdea {
 	id: string;
@@ -122,7 +123,7 @@ export const useAIStore = create<AIState>()((set) => ({
 			savedIdeas: [
 				...state.savedIdeas,
 				{
-					id: crypto.randomUUID(),
+					id: generateUUID(),
 					content,
 					savedAt: Date.now(),
 				},
diff --git a/apps/web/src/types/timeline.ts b/apps/web/src/types/timeline.ts
index 78904ad3c7..cb3b5e650a 100644
--- a/apps/web/src/types/timeline.ts
+++ b/apps/web/src/types/timeline.ts
@@ -1,6 +1,7 @@
 import type { ElementAnimations } from "./animation";
 import type { Effect, EffectParamValues } from "./effects";
 import type { BlendMode, Transform, CropRect, MaskShape } from "./rendering";
+import type { TrackAudioEffect } from "@/lib/audio/audio-effects";
 
 export interface TransitionData {
 	type: string;
@@ -62,6 +63,7 @@ export interface VideoTrack extends BaseTrack {
 	hidden: boolean;
 	volume?: number;
 	solo?: boolean;
+	audioEffects?: TrackAudioEffect[];
 }
 
 export interface TextTrack extends BaseTrack {
@@ -77,6 +79,7 @@ export interface AudioTrack extends BaseTrack {
 	volume?: number;
 	pan?: number;
 	solo?: boolean;
+	audioEffects?: TrackAudioEffect[];
 }
 
 export interface StickerTrack extends BaseTrack {
diff --git a/apps/web/src/utils/id.ts b/apps/web/src/utils/id.ts
index e452af434e..229bf6ecac 100644
--- a/apps/web/src/utils/id.ts
+++ b/apps/web/src/utils/id.ts
@@ -1,3 +1,19 @@
+function fillRandomBytes(bytes: Uint8Array): void {
+	if (
+		typeof crypto !== "undefined" &&
+		typeof crypto.getRandomValues === "function"
+	) {
+		crypto.getRandomValues(bytes);
+		return;
+	}
+
+	// Last-resort fallback for non-secure / legacy contexts.
+	for (let i = 0; i < bytes.length; i++) {
+		bytes[i] = Math.floor(Math.random() * 256);
+	}
+}
+
+/** UUID v4 — works in insecure HTTP contexts where crypto.randomUUID is missing. */
 export function generateUUID(): string {
 	if (
 		typeof crypto !== "undefined" &&
@@ -7,7 +23,7 @@ export function generateUUID(): string {
 	}
 
 	const bytes = new Uint8Array(16);
-	crypto.getRandomValues(bytes);
+	fillRandomBytes(bytes);
 
 	bytes[6] = (bytes[6] & 0x0f) | 0x40;
 	bytes[8] = (bytes[8] & 0x3f) | 0x80;
diff --git a/docker-compose.yml b/docker-compose.yml
index 5e548bce4f..f8df6ca356 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -135,11 +135,11 @@ services:
     volumes:
       - tts_models:/root/.cache
     environment:
-      - TTS_AUTOLOAD=true
+      - TTS_AUTOLOAD=false
     deploy:
       resources:
         limits:
-          memory: 6g
+          memory: 3g
     healthcheck:
       test: ["CMD-SHELL", "curl -f http://localhost:8422/health || exit 1"]
       interval: 30s
@@ -261,6 +261,13 @@ services:
       args:
         - FREESOUND_CLIENT_ID=${FREESOUND_CLIENT_ID}
         - FREESOUND_API_KEY=${FREESOUND_API_KEY}
+        - NEXT_PUBLIC_SITE_URL=${NEXT_PUBLIC_SITE_URL:-http://localhost:3100}
+        - NEXT_PUBLIC_MARBLE_API_URL=${NEXT_PUBLIC_MARBLE_API_URL:-https://api.marblecms.com}
+        - NEXT_PUBLIC_AI_BACKEND_URL=${NEXT_PUBLIC_AI_BACKEND_URL:-http://localhost:8420}
+        - NEXT_PUBLIC_OLLAMA_URL=${NEXT_PUBLIC_OLLAMA_URL:-http://localhost:11434}
+        - NEXT_PUBLIC_WHISPER_SERVICE_URL=${NEXT_PUBLIC_WHISPER_SERVICE_URL:-http://localhost:8421}
+        - NEXT_PUBLIC_TTS_SERVICE_URL=${NEXT_PUBLIC_TTS_SERVICE_URL:-http://localhost:8422}
+        - NEXT_PUBLIC_IMAGE_SERVICE_URL=${NEXT_PUBLIC_IMAGE_SERVICE_URL:-http://localhost:8423}
     restart: unless-stopped
     ports:
       - "3100:3000"
@@ -270,9 +277,10 @@ services:
       - BETTER_AUTH_SECRET=your-production-secret-key-here
       - UPSTASH_REDIS_REST_URL=http://serverless-redis-http:80
       - UPSTASH_REDIS_REST_TOKEN=example_token
-      - NEXT_PUBLIC_SITE_URL=http://localhost:3100
-      - NEXT_PUBLIC_MARBLE_API_URL=https://api.marblecms.com
-      - NEXT_PUBLIC_AI_BACKEND_URL=http://localhost:8420
+      - NEXT_PUBLIC_SITE_URL=${NEXT_PUBLIC_SITE_URL:-http://localhost:3100}
+      - NEXT_PUBLIC_MARBLE_API_URL=${NEXT_PUBLIC_MARBLE_API_URL:-https://api.marblecms.com}
+      - NEXT_PUBLIC_AI_BACKEND_URL=${NEXT_PUBLIC_AI_BACKEND_URL:-http://localhost:8420}
+      - NEXT_PUBLIC_OLLAMA_URL=${NEXT_PUBLIC_OLLAMA_URL:-http://localhost:11434}
       - MARBLE_WORKSPACE_KEY=${MARBLE_WORKSPACE_KEY:-placeholder}
       - FREESOUND_CLIENT_ID=${FREESOUND_CLIENT_ID}
       - FREESOUND_API_KEY=${FREESOUND_API_KEY}
@@ -282,9 +290,9 @@ services:
       - R2_SECRET_ACCESS_KEY=${R2_SECRET_ACCESS_KEY:-placeholder}
       - R2_BUCKET_NAME=${R2_BUCKET_NAME:-opencut-transcription}
       - MODAL_TRANSCRIPTION_URL=${MODAL_TRANSCRIPTION_URL:-http://localhost:0}
-      - NEXT_PUBLIC_WHISPER_SERVICE_URL=http://localhost:8421
-      - NEXT_PUBLIC_TTS_SERVICE_URL=http://localhost:8422
-      - NEXT_PUBLIC_IMAGE_SERVICE_URL=http://localhost:8423
+      - NEXT_PUBLIC_WHISPER_SERVICE_URL=${NEXT_PUBLIC_WHISPER_SERVICE_URL:-http://localhost:8421}
+      - NEXT_PUBLIC_TTS_SERVICE_URL=${NEXT_PUBLIC_TTS_SERVICE_URL:-http://localhost:8422}
+      - NEXT_PUBLIC_IMAGE_SERVICE_URL=${NEXT_PUBLIC_IMAGE_SERVICE_URL:-http://localhost:8423}
     depends_on:
       db:
         condition: service_healthy
@@ -293,7 +301,7 @@ services:
       ai-backend:
         condition: service_healthy
     healthcheck:
-      test: ["CMD-SHELL", "curl -f http://localhost:3000/api/health || exit 1"]
+      test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:3000/api/health').then(r=>process.exit(r.ok?0:1)).catch(()=>process.exit(1))"]
       interval: 30s
       timeout: 10s
       retries: 5
diff --git a/scripts/do-cloud-init.yaml b/scripts/do-cloud-init.yaml
new file mode 100644
index 0000000000..56600061ba
--- /dev/null
+++ b/scripts/do-cloud-init.yaml
@@ -0,0 +1,128 @@
+#cloud-config
+package_update: true
+package_upgrade: false
+
+packages:
+  - git
+  - curl
+  - ca-certificates
+  - gnupg
+  - jq
+  - ufw
+  - python3
+  - openssl
+
+write_files:
+  - path: /opt/opencut/deploy.sh
+    permissions: "0755"
+    content: |
+      #!/usr/bin/env bash
+      set -euo pipefail
+      exec > >(tee -a /var/log/opencut-deploy.log) 2>&1
+
+      echo "[opencut] deploy started $(date -u +%Y-%m-%dT%H:%M:%SZ)"
+
+      PUBLIC_IP="$(curl -fsS http://169.254.169.254/metadata/v1/interfaces/public/0/ipv4/address)"
+      AUTH_SECRET="$(openssl rand -hex 32)"
+      APP_DIR=/opt/opencut/OpenCut-AI
+
+      if ! command -v docker >/dev/null 2>&1; then
+        curl -fsSL https://get.docker.com | sh
+        systemctl enable --now docker
+      fi
+
+      mkdir -p /opt/opencut
+      if [ ! -d "$APP_DIR/.git" ]; then
+        git clone --depth 1 https://github.com/Ekaanth/OpenCut-AI.git "$APP_DIR"
+      fi
+      cd "$APP_DIR"
+
+      # Bake public URLs into the Next.js build (NEXT_PUBLIC_* is compile-time)
+      sed -i \
+        -e "s|ENV NEXT_PUBLIC_SITE_URL=\"http://localhost:3000\"|ENV NEXT_PUBLIC_SITE_URL=\"http://${PUBLIC_IP}:3100\"|" \
+        -e "s|ENV NEXT_PUBLIC_AI_BACKEND_URL=\"http://localhost:8420\"|ENV NEXT_PUBLIC_AI_BACKEND_URL=\"http://${PUBLIC_IP}:8420\"|" \
+        apps/web/Dockerfile
+
+      cat > .env <<EOF
+      FREESOUND_CLIENT_ID=
+      FREESOUND_API_KEY=
+      MARBLE_WORKSPACE_KEY=placeholder
+      OLLAMA_DEFAULT_MODEL=llama3.2:1b
+      WHISPER_MODEL_SIZE=base
+      AI_MEMORY_BUDGET=auto
+      AI_MODEL_TIER=auto
+      EOF
+
+      PUBLIC_IP="$PUBLIC_IP" AUTH_SECRET="$AUTH_SECRET" python3 - <<'PY'
+      from pathlib import Path
+      import os
+
+      ip = os.environ["PUBLIC_IP"]
+      secret = os.environ["AUTH_SECRET"]
+      p = Path("docker-compose.yml")
+      text = p.read_text()
+      replacements = [
+          (
+              "BETTER_AUTH_SECRET=your-production-secret-key-here",
+              f"BETTER_AUTH_SECRET={secret}",
+          ),
+          (
+              "NEXT_PUBLIC_SITE_URL=http://localhost:3100",
+              f"NEXT_PUBLIC_SITE_URL=http://{ip}:3100",
+          ),
+          (
+              'OPENCUTAI_CORS_ORIGINS=["http://localhost:3000","http://localhost:3100","http://localhost:5173"]',
+              f'OPENCUTAI_CORS_ORIGINS=["http://{ip}:3100","http://localhost:3000","http://localhost:3100"]',
+          ),
+          (
+              "NEXT_PUBLIC_AI_BACKEND_URL=http://localhost:8420",
+              f"NEXT_PUBLIC_AI_BACKEND_URL=http://{ip}:8420",
+          ),
+          (
+              "NEXT_PUBLIC_WHISPER_SERVICE_URL=http://localhost:8421",
+              f"NEXT_PUBLIC_WHISPER_SERVICE_URL=http://{ip}:8421",
+          ),
+          (
+              "NEXT_PUBLIC_TTS_SERVICE_URL=http://localhost:8422",
+              f"NEXT_PUBLIC_TTS_SERVICE_URL=http://{ip}:8422",
+          ),
+          (
+              "NEXT_PUBLIC_IMAGE_SERVICE_URL=http://localhost:8423",
+              f"NEXT_PUBLIC_IMAGE_SERVICE_URL=http://{ip}:8423",
+          ),
+      ]
+      for old, new in replacements:
+          if old not in text:
+              raise SystemExit(f"patch target missing: {old}")
+          text = text.replace(old, new)
+      p.write_text(text)
+      print(f"[opencut] patched compose for http://{ip}:3100")
+      PY
+
+      ufw allow OpenSSH
+      ufw allow 3100/tcp
+      ufw allow 8420:8426/tcp
+      ufw allow 8430/tcp
+      ufw --force enable || true
+
+      export DOCKER_BUILDKIT=1
+      export COMPOSE_DOCKER_CLI_BUILD=1
+
+      echo "[opencut] building images (this can take a long time)..."
+      docker compose build --parallel
+
+      echo "[opencut] starting stack..."
+      docker compose up -d
+
+      sleep 20
+      docker compose exec -T ollama ollama pull llama3.2:1b || true
+
+      echo "[opencut] status:"
+      docker compose ps
+      echo "[opencut] READY at http://${PUBLIC_IP}:3100"
+      echo "[opencut] deploy finished $(date -u +%Y-%m-%dT%H:%M:%SZ)"
+
+runcmd:
+  - [bash, -lc, "/opt/opencut/deploy.sh"]
+
+final_message: "OpenCut AI cloud-init finished scheduling deploy.sh"
diff --git a/services/ai-backend/app/models/__init__.py b/services/ai-backend/app/models/__init__.py
new file mode 100644
index 0000000000..981f1f49ba
--- /dev/null
+++ b/services/ai-backend/app/models/__init__.py
@@ -0,0 +1,53 @@
+"""Pydantic request/response models for the AI backend."""
+
+from app.models.audio import TTSRequest
+from app.models.command import CommandRequest, CommandResponse, EditorAction
+from app.models.engagement import (
+    AudioSyncScore,
+    CuriosityScore,
+    EmotionalArcScore,
+    EngagementScore,
+    EnhancementSuggestion,
+    EnergyScore,
+    FacePresenceScore,
+    HookScore,
+    JobStatus,
+    ScoreBatchRequest,
+    ScoreClipRequest,
+    ScoredClip,
+    ViralityScore,
+    YouTubeVideoMeta,
+)
+from app.models.generation import EnhancePromptRequest, ImageGenParams, InfographicRequest
+from app.models.transcription import (
+    TranscriptionResult,
+    TranscriptionSegment,
+    TranscriptionWord,
+)
+
+__all__ = [
+    "TTSRequest",
+    "CommandRequest",
+    "CommandResponse",
+    "EditorAction",
+    "AudioSyncScore",
+    "CuriosityScore",
+    "EmotionalArcScore",
+    "EngagementScore",
+    "EnhancementSuggestion",
+    "EnergyScore",
+    "FacePresenceScore",
+    "HookScore",
+    "JobStatus",
+    "ScoreBatchRequest",
+    "ScoreClipRequest",
+    "ScoredClip",
+    "ViralityScore",
+    "YouTubeVideoMeta",
+    "EnhancePromptRequest",
+    "ImageGenParams",
+    "InfographicRequest",
+    "TranscriptionResult",
+    "TranscriptionSegment",
+    "TranscriptionWord",
+]
diff --git a/services/ai-backend/app/models/audio.py b/services/ai-backend/app/models/audio.py
new file mode 100644
index 0000000000..d969c05df2
--- /dev/null
+++ b/services/ai-backend/app/models/audio.py
@@ -0,0 +1,15 @@
+"""Audio / TTS request models."""
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class TTSRequest(BaseModel):
+    """Text-to-speech generation request (proxied to tts-service)."""
+
+    model_config = ConfigDict(populate_by_name=True)
+
+    text: str
+    language: str = "en"
+    speaker_wav: str | None = Field(default=None, alias="speakerWav")
+    speaker: str | None = None
+    speed: float = 1.0
diff --git a/services/ai-backend/app/models/command.py b/services/ai-backend/app/models/command.py
new file mode 100644
index 0000000000..b91d1a1f95
--- /dev/null
+++ b/services/ai-backend/app/models/command.py
@@ -0,0 +1,33 @@
+"""LLM command panel request/response models."""
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class EditorAction(BaseModel):
+    """A single structured editor action produced by the LLM."""
+
+    type: str
+    target: str | None = None
+    params: dict = Field(default_factory=dict)
+    description: str = ""
+
+
+class CommandRequest(BaseModel):
+    """Natural-language editing command from the frontend."""
+
+    model_config = ConfigDict(populate_by_name=True)
+
+    command: str
+    timeline_state: dict | None = Field(default=None, alias="timelineState")
+    model: str | None = None
+
+
+class CommandResponse(BaseModel):
+    """Structured response for an interpreted editing command."""
+
+    actions: list[EditorAction] = Field(default_factory=list)
+    explanation: str = ""
+    confidence: float = 0.5
+    raw_response: str | None = None
+    needs_clarification: bool = False
+    clarification_question: str | None = None
diff --git a/services/ai-backend/app/models/engagement.py b/services/ai-backend/app/models/engagement.py
new file mode 100644
index 0000000000..7541aea5a7
--- /dev/null
+++ b/services/ai-backend/app/models/engagement.py
@@ -0,0 +1,222 @@
+"""Engagement scoring, YouTube ingest, and clip models."""
+
+from pydantic import BaseModel, Field, computed_field
+
+
+# ── Sub-scores ────────────────────────────────────────────────────────
+
+
+class HookScore(BaseModel):
+    """Hook strength of the first ~3 seconds."""
+
+    visual_novelty: float = 0.0
+    audio_energy_spike: float = 0.0
+    early_face_present: bool = False
+    hook_type: str = "neutral"
+    hook_type_confidence: float = 0.0
+    speech_rate: float = 0.0
+    composite: float = 50.0
+
+
+class CuriosityScore(BaseModel):
+    """Curiosity-gap signals in the transcript."""
+
+    has_question: bool = False
+    has_bold_claim: bool = False
+    has_open_loop: bool = False
+    gap_count: int = 0
+    composite: float = 50.0
+
+
+class EnergyScore(BaseModel):
+    """Audio energy dynamics."""
+
+    mean_energy: float = 0.0
+    peak_energy: float = 0.0
+    energy_variance: float = 0.0
+    has_dynamic_range: bool = False
+    composite: float = 50.0
+
+
+class AudioSyncScore(BaseModel):
+    """Beat / caption sync quality."""
+
+    bpm: float | None = None
+    beat_count: int = 0
+    caption_beat_alignment: float = 0.0
+    composite: float = 50.0
+
+
+class FacePresenceScore(BaseModel):
+    """Face presence ratio vs optimal 30–40% target."""
+
+    face_ratio: float = 0.0
+    is_optimal: bool = False
+    early_face_present: bool = False
+    composite: float = 50.0
+
+
+class EmotionalArcScore(BaseModel):
+    """Emotional arc / pacing structure."""
+
+    has_strong_open: bool = False
+    has_buildup: bool = False
+    has_peak: bool = False
+    peak_timestamp: float = 0.0
+    dominant_emotion: str = "calm"
+    composite: float = 50.0
+
+
+class ViralityScore(BaseModel):
+    """LLM-predicted viral potential."""
+
+    hook_strength: int = 0
+    shareability: int = 0
+    emotional_impact: int = 0
+    standalone_value: int = 0
+    reason: str = ""
+    suggested_title: str = ""
+    composite: float = 50.0
+
+
+class EnhancementSuggestion(BaseModel):
+    """Actionable improvement suggestion for a weak engagement signal."""
+
+    signal: str
+    current_score: float
+    suggestion: str
+    action_type: str = "manual"
+    expected_impact: str = "medium"
+
+
+# ── Aggregate score ───────────────────────────────────────────────────
+
+
+def _grade_for_score(composite: float) -> tuple[str, str]:
+    """Map a 0–100 composite score to (grade letter, label)."""
+    if composite >= 85:
+        return "A", "Excellent"
+    if composite >= 70:
+        return "B", "Strong"
+    if composite >= 50:
+        return "C", "Average"
+    if composite >= 35:
+        return "D", "Below average"
+    return "F", "Needs work"
+
+
+class EngagementScore(BaseModel):
+    """Full engagement breakdown with composite score and suggestions."""
+
+    hook: HookScore = Field(default_factory=HookScore)
+    curiosity: CuriosityScore = Field(default_factory=CuriosityScore)
+    energy: EnergyScore = Field(default_factory=EnergyScore)
+    audio_sync: AudioSyncScore = Field(default_factory=AudioSyncScore)
+    face_presence: FacePresenceScore = Field(default_factory=FacePresenceScore)
+    emotional_arc: EmotionalArcScore = Field(default_factory=EmotionalArcScore)
+    virality: ViralityScore = Field(default_factory=ViralityScore)
+    suggestions: list[EnhancementSuggestion] = Field(default_factory=list)
+
+    def compute_composite(self) -> float:
+        """Weighted composite using configured engagement weights."""
+        from app.config import settings
+
+        return (
+            self.hook.composite * settings.ENGAGEMENT_HOOK_WEIGHT
+            + self.curiosity.composite * settings.ENGAGEMENT_CURIOSITY_WEIGHT
+            + self.virality.composite * settings.ENGAGEMENT_VIRALITY_WEIGHT
+            + self.energy.composite * settings.ENGAGEMENT_ENERGY_WEIGHT
+            + self.emotional_arc.composite * settings.ENGAGEMENT_EMOTION_WEIGHT
+            + self.audio_sync.composite * settings.ENGAGEMENT_AUDIO_SYNC_WEIGHT
+            + self.face_presence.composite * settings.ENGAGEMENT_FACE_WEIGHT
+        )
+
+    def to_response(self) -> dict:
+        """Serialize to the API shape expected by the web client."""
+        composite = round(min(100.0, max(0.0, self.compute_composite())), 1)
+        grade, grade_label = _grade_for_score(composite)
+        return {
+            "hook": self.hook.model_dump(),
+            "curiosity": self.curiosity.model_dump(),
+            "energy": self.energy.model_dump(),
+            "audio_sync": self.audio_sync.model_dump(),
+            "face_presence": self.face_presence.model_dump(),
+            "emotional_arc": self.emotional_arc.model_dump(),
+            "virality": self.virality.model_dump(),
+            "suggestions": [s.model_dump() for s in self.suggestions],
+            "composite": composite,
+            "grade": grade,
+            "grade_label": grade_label,
+        }
+
+    @property
+    def composite(self) -> float:
+        return self.compute_composite()
+
+
+# ── Request models ────────────────────────────────────────────────────
+
+
+class ScoreClipRequest(BaseModel):
+    """Score a single clip from transcript / audio / video paths."""
+
+    audio_path: str | None = None
+    video_path: str | None = None
+    transcript_text: str = ""
+    transcript_segments: list[dict] | None = None
+    start: float = 0.0
+    end: float = 0.0
+    title: str | None = None
+
+
+class ScoreBatchRequest(BaseModel):
+    """Batch scoring request for multiple clips."""
+
+    clips: list[ScoreClipRequest] = Field(default_factory=list)
+
+
+class ScoredClip(BaseModel):
+    """A detected clip with engagement score attached."""
+
+    index: int = 0
+    title: str = ""
+    start: float = 0.0
+    end: float = 0.0
+    transcript_preview: str = ""
+    tags: list[str] = Field(default_factory=list)
+    engagement: EngagementScore = Field(default_factory=EngagementScore)
+
+    @computed_field  # type: ignore[prop-decorator]
+    @property
+    def duration(self) -> float:
+        return max(0.0, self.end - self.start)
+
+
+# ── YouTube / jobs ────────────────────────────────────────────────────
+
+
+class YouTubeVideoMeta(BaseModel):
+    """Metadata for an ingested YouTube video."""
+
+    video_id: str
+    title: str = "Untitled"
+    channel_name: str = "Unknown"
+    channel_id: str = ""
+    duration_seconds: int = 0
+    thumbnail_url: str = ""
+    upload_date: str = ""
+    view_count: int | None = None
+    is_live: bool = False
+    is_private: bool = False
+    warning: str | None = None
+
+
+class JobStatus(BaseModel):
+    """Background job status for YouTube / clip pipelines."""
+
+    job_id: str
+    status: str = "pending"
+    progress: float = 0.0
+    message: str = ""
+    result: dict | None = None
+    error: str | None = None
diff --git a/services/ai-backend/app/models/generation.py b/services/ai-backend/app/models/generation.py
new file mode 100644
index 0000000000..e852f31ddb
--- /dev/null
+++ b/services/ai-backend/app/models/generation.py
@@ -0,0 +1,36 @@
+"""Image generation and infographic request models."""
+
+from pydantic import BaseModel, ConfigDict, Field
+
+
+class ImageGenParams(BaseModel):
+    """Parameters for text-to-image generation (proxied to image-service)."""
+
+    model_config = ConfigDict(populate_by_name=True)
+
+    prompt: str
+    negative_prompt: str = Field(default="", alias="negativePrompt")
+    width: int = 512
+    height: int = 512
+    steps: int = 20
+    guidance_scale: float = Field(default=7.5, alias="guidanceScale")
+    seed: int | None = None
+    model: str | None = None
+
+
+class EnhancePromptRequest(BaseModel):
+    """Request to expand a short prompt into a detailed diffusion prompt."""
+
+    prompt: str
+    style: str = "photorealistic"
+
+
+class InfographicRequest(BaseModel):
+    """Request to render a simple infographic overlay PNG."""
+
+    topic: str
+    data_points: list[dict] = Field(default_factory=list)
+    style: str = "modern"
+    width: int = 1080
+    height: int = 1080
+    background_color: tuple[int, int, int, int] | str = (0, 0, 0, 0)
diff --git a/services/ai-backend/app/models/transcription.py b/services/ai-backend/app/models/transcription.py
new file mode 100644
index 0000000000..ff18d0e754
--- /dev/null
+++ b/services/ai-backend/app/models/transcription.py
@@ -0,0 +1,34 @@
+"""Whisper transcription result models."""
+
+from pydantic import BaseModel, Field
+
+
+class TranscriptionWord(BaseModel):
+    """Word-level timestamp from faster-whisper."""
+
+    word: str
+    start: float
+    end: float
+    probability: float = 0.0
+
+
+class TranscriptionSegment(BaseModel):
+    """A transcribed segment with optional word timings."""
+
+    id: int
+    text: str
+    start: float
+    end: float
+    words: list[TranscriptionWord] = Field(default_factory=list)
+    avg_logprob: float = 0.0
+    no_speech_prob: float = 0.0
+    speaker: str | None = None
+
+
+class TranscriptionResult(BaseModel):
+    """Full transcription output."""
+
+    text: str = ""
+    segments: list[TranscriptionSegment] = Field(default_factory=list)
+    language: str = ""
+    duration: float = 0.0
diff --git a/services/ai-backend/app/routes/engagement.py b/services/ai-backend/app/routes/engagement.py
index 5666bcda9d..22d0e79406 100644
--- a/services/ai-backend/app/routes/engagement.py
+++ b/services/ai-backend/app/routes/engagement.py
@@ -9,7 +9,7 @@
 import uuid
 
 from fastapi import APIRouter, File, Form, HTTPException, UploadFile
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 
 from app.config import settings
 from app.models.engagement import (
diff --git a/services/tts-service/requirements.lock b/services/tts-service/requirements.lock
index 16957e10d9..4f883e1f8b 100644
--- a/services/tts-service/requirements.lock
+++ b/services/tts-service/requirements.lock
@@ -408,7 +408,7 @@ torch==2.5.1
     #   coqui-tts
     #   coqui-tts-trainer
     #   encodec
-torchaudio==2.11.0
+torchaudio==2.5.1
     # via
     #   coqui-tts
     #   encodec
diff --git a/services/tts-service/requirements.txt b/services/tts-service/requirements.txt
index d38e60f6c9..f52c470515 100644
--- a/services/tts-service/requirements.txt
+++ b/services/tts-service/requirements.txt
@@ -3,4 +3,5 @@ uvicorn[standard]==0.30.0
 python-multipart==0.0.9
 aiofiles==24.1.0
 torch>=2.1.0,<2.6.0
+torchaudio==2.5.1
 coqui-tts==0.24.2