diff --git a/apps/docs/api/ai.md b/apps/docs/api/ai.md
index b9cb1a77..ee4f238d 100644
--- a/apps/docs/api/ai.md
+++ b/apps/docs/api/ai.md
@@ -6,119 +6,173 @@ description: AI engine reference with all local ML tools. Background removal, up
 
 The `@snapotter/ai` package bridges Node.js to a **persistent Python sidecar** for all ML operations. The dispatcher process stays alive between requests for fast warm-start performance. GPU is auto-detected at startup and used when available.
 
-19 AI tool routes. All models run locally - no internet required after initial model download.
+19 Python sidecar AI tools across four modalities (image, audio, video, document), plus 2 tools with optional AI capabilities. All models run locally -- no internet required after initial model download.
 
 ## Architecture
 
 ```
 Node.js Tool Route
-      │
-      ▼
+      |
+      v
  @snapotter/ai bridge.ts
-      │ (stdin/stdout JSON + stderr progress events)
-      ▼
- Python dispatcher (persistent process)
-      │
-      ├─ remove_bg.py        (rembg / BiRefNet)
-      ├─ upscale.py          (RealESRGAN)
-      ├─ inpaint.py          (LaMa ONNX)
-      ├─ ocr.py              (PaddleOCR / Tesseract)
-      ├─ detect_faces.py     (MediaPipe)
-      ├─ face_landmarks.py   (MediaPipe landmarks)
-      ├─ enhance_faces.py    (GFPGAN / CodeFormer)
-      ├─ colorize.py         (DDColor)
-      ├─ noise_removal.py    (tiered denoising)
-      ├─ red_eye_removal.py  (landmark + color analysis)
-      ├─ restore.py          (scratch repair + enhancement + denoising)
-      ├─ transparency_fix.py (BiRefNet HR-matting + defringe)
-      └─ seam_carving        (Go caire binary - not Python)
+      | (stdin/stdout JSON + stderr progress events)
+      v
+ Python dispatcher (persistent process, "ai" profile)
+      |
+      |-- remove_bg.py        (rembg / BiRefNet)
+      |-- upscale.py          (RealESRGAN)
+      |-- inpaint.py          (LaMa ONNX)
+      |-- outpaint.py         (LaMa canvas expansion)
+      |-- ocr.py              (PaddleOCR / Tesseract)
+      |-- ocr_pdf.py          (page-by-page document OCR)
+      |-- ocr_preprocess.py   (image enhancement for OCR)
+      |-- detect_faces.py     (MediaPipe)
+      |-- face_landmarks.py   (MediaPipe landmarks)
+      |-- enhance_faces.py    (GFPGAN / CodeFormer)
+      |-- colorize.py         (DDColor)
+      |-- noise_removal.py    (SCUNet / tiered denoising)
+      |-- red_eye_removal.py  (landmark + color analysis)
+      |-- restore.py          (scratch repair + enhancement + denoising)
+      |-- transcribe.py       (faster-whisper speech-to-text)
+      +-- install_feature.py  (on-demand bundle installer)
 ```
 
+A separate "docs" dispatcher profile replaces the AI allowlist with document-processing scripts (`doc_pagecount`, `doc_health`, `doc_flatten`, `doc_redact`, `doc_text`, `doc_to_word`, `doc_metadata`, `doc_html_pdf`) and skips heavy ML imports.
+
 **Timeouts:** 300 s default; OCR and BiRefNet background removal get 600 s.
 
+## Feature Bundles
+
+Each AI tool requires a model bundle to be installed before use. Bundles are installed on demand via the admin UI or `install_feature.py`.
+
+| Bundle | Size | Tools |
+|--------|------|-------|
+| `background-removal` | 4-5 GB | remove-background, passport-photo, transparency-fixer, background-replace, blur-background |
+| `face-detection` | 200-300 MB | blur-faces, red-eye-removal, smart-crop |
+| `object-eraser-colorize` | 1-2 GB | erase-object, colorize, ai-canvas-expand |
+| `upscale-enhance` | 4-5 GB | upscale, enhance-faces, noise-removal |
+| `photo-restoration` | 800 MB - 1 GB | restore-photo |
+| `ocr` | 3-4 GB | ocr, ocr-pdf |
+| `transcription` | ~600 MB | transcribe-audio, auto-subtitles |
+
+---
+
 ## Background Removal
 
-**Function:** `removeBackground`  
 **Tool route:** `remove-background`  
 **Model:** rembg with BiRefNet (default) or U2-Net variants
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `model` | string | `birefnet-general` | Model variant - see table below |
-| `alphaMattingForeground` | number (1–255) | 240 | Foreground threshold for alpha matting |
-| `alphaMattingBackground` | number (1–255) | 10 | Background threshold for alpha matting |
-| `returnMask` | boolean | false | Return the mask instead of the cutout |
-| `backgroundColor` | string | - | Fill removed area (hex color or "transparent") |
+| `model` | string | - | Model variant (optional override) |
+| `backgroundType` | string | `"transparent"` | One of: `transparent`, `color`, `gradient`, `blur`, `image` |
+| `backgroundColor` | string | - | Hex color for solid background |
+| `gradientColor1` | string | - | First gradient color |
+| `gradientColor2` | string | - | Second gradient color |
+| `gradientAngle` | number | - | Gradient angle in degrees |
+| `blurEnabled` | boolean | - | Enable background blur effect |
+| `blurIntensity` | number (0-100) | - | Blur intensity |
+| `shadowEnabled` | boolean | - | Enable drop shadow on subject |
+| `shadowOpacity` | number (0-100) | - | Shadow opacity |
+| `outputFormat` | string | - | Output format: `png`, `webp`, or `avif` |
+| `edgeRefine` | integer (0-3) | - | Edge refinement level |
+| `decontaminate` | boolean | - | Remove color bleed from edges |
 
-**Available models:**
+## Background Replace
 
-| Model ID | Best for |
-|----------|---------|
-| `birefnet-general` | General purpose (default) |
-| `birefnet-portrait` | People / portraits |
-| `birefnet-dis` | Dichotomous Image Segmentation |
-| `birefnet-hrsod` | High-resolution salient objects |
-| `birefnet-cod` | Camouflaged objects |
-| `u2net` | Fast general purpose |
-| `u2net_human_seg` | Human segmentation |
-| `isnet-general-use` | High quality general |
+**Tool route:** `background-replace`  
+**Model:** rembg / BiRefNet (shared with remove-background)
 
-## Image Upscaling
-
-**Function:** `upscale`  
-**Tool route:** `upscale`  
-**Model:** RealESRGAN (with Lanczos fallback on CPU-constrained systems)
+Removes the background and replaces it with a solid color or gradient.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `scale` | 2 \| 4 | 4 | Upscale factor |
-| `model` | string | `realesrgan-x4plus` | Model variant |
-| `faceEnhance` | boolean | false | Apply GFPGAN face enhancement pass |
-| `denoise` | number (0–1) | 0.5 | Denoising strength |
-| `format` | string | - | Output format override |
-| `quality` | number | 95 | Output quality (for JPEG/WebP) |
+| `backgroundType` | `"color"` \| `"gradient"` | `"color"` | Background mode |
+| `color` | string | `"#ffffff"` | Background hex color (when `backgroundType` is `color`) |
+| `gradientColor1` | string | - | First gradient hex color |
+| `gradientColor2` | string | - | Second gradient hex color |
+| `gradientAngle` | integer (0-360) | `180` | Gradient angle in degrees |
+| `feather` | integer (0-20) | `0` | Edge feathering radius |
+| `format` | `"png"` \| `"webp"` | `"png"` | Output format |
+
+## Blur Background
+
+**Tool route:** `blur-background`  
+**Model:** rembg / BiRefNet (shared with remove-background)
+
+Blurs the background while keeping the subject sharp.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `intensity` | integer (1-100) | `50` | Blur intensity |
+| `feather` | integer (0-20) | `0` | Edge feathering radius |
+| `format` | `"png"` \| `"webp"` | `"png"` | Output format |
+
+## Image Upscaling
+
+**Tool route:** `upscale`  
+**Model:** RealESRGAN (with Lanczos fallback when unavailable)
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `scale` | number | `2` | Upscale factor |
+| `model` | string | `"auto"` | Model variant |
+| `faceEnhance` | boolean | `false` | Apply GFPGAN face enhancement pass |
+| `denoise` | number | `0` | Denoising strength |
+| `format` | string | `"auto"` | Output format override |
+| `quality` | number | `95` | Output quality (1-100) |
 
 ## OCR / Text Extraction
 
-**Function:** `extractText`  
 **Tool route:** `ocr`  
 **Models:** Tesseract (fast), PaddleOCR PP-OCRv5 (balanced), PaddleOCR-VL 1.5 (best)
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `quality` | `fast` \| `balanced` \| `best` | `balanced` | Processing tier |
-| `language` | string | `en` | Language code (ISO 639-1) |
-| `enhance` | boolean | false | Pre-process image to improve OCR accuracy |
+| `quality` | `"fast"` \| `"balanced"` \| `"best"` | `"balanced"` | Processing tier |
+| `language` | string | `"auto"` | Language: `auto`, `en`, `de`, `fr`, `es`, `zh`, `ja`, `ko` |
+| `enhance` | boolean | `true` | Pre-process image to improve OCR accuracy |
+| `engine` | string | - | Deprecated. Maps `tesseract` to `fast`, `paddleocr` to `balanced` |
 
 Returns structured results with bounding boxes, confidence scores, and extracted text blocks.
 
+## PDF OCR
+
+**Tool route:** `ocr-pdf`  
+**Models:** Same tier system as image OCR
+
+Extracts text from scanned PDF documents using AI-powered OCR, page by page.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `quality` | `"fast"` \| `"balanced"` \| `"best"` | `"balanced"` | Processing tier |
+| `language` | string | `"auto"` | Language: `auto`, `en`, `de`, `fr`, `es`, `zh`, `ja`, `ko` |
+| `pages` | string | `"all"` | Page selection: `"all"`, `"1-3"`, `"1,3,5"` |
+
 ## Face / PII Blur
 
-**Function:** `blurFaces`  
 **Tool route:** `blur-faces`  
 **Model:** MediaPipe face detection
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `blurRadius` | number | 30 | Gaussian blur radius |
-| `sensitivity` | number (0–1) | 0.5 | Detection confidence threshold |
+| `blurRadius` | number (1-100) | `30` | Gaussian blur radius |
+| `sensitivity` | number (0-1) | `0.5` | Detection confidence threshold |
 
 ## Face Enhancement
 
-**Function:** `enhanceFaces`  
 **Tool route:** `enhance-faces`  
 **Models:** GFPGAN, CodeFormer
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `model` | `gfpgan` \| `codeformer` | `gfpgan` | Enhancement model |
-| `strength` | number (0–1) | 0.7 | Enhancement strength |
-| `sensitivity` | number (0–1) | 0.5 | Face detection threshold |
-| `centerFace` | boolean | false | Focus enhancement on center face only |
+| `model` | `"auto"` \| `"gfpgan"` \| `"codeformer"` | `"auto"` | Enhancement model |
+| `strength` | number (0-1) | `0.8` | Enhancement strength |
+| `sensitivity` | number (0-1) | `0.5` | Face detection threshold |
+| `onlyCenterFace` | boolean | `false` | Enhance only the most central face |
 
 ## AI Colorization
 
-**Function:** `colorize`  
 **Tool route:** `colorize`  
 **Model:** DDColor (with OpenCV DNN fallback)
 
@@ -126,144 +180,182 @@ Converts black-and-white or grayscale photos to full color.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `intensity` | number (0–1) | 0.85 | Color saturation strength |
-| `model` | string | `ddcolor` | Model variant |
+| `intensity` | number (0-1) | `1.0` | Color saturation strength |
+| `model` | `"auto"` \| `"ddcolor"` \| `"opencv"` | `"auto"` | Model variant |
 
 ## Noise Removal
 
-**Function:** `noiseRemoval`  
-**Tool route:** `noise-removal`
-
-Three-tier denoising pipeline (fast: OpenCV bilateral filter; balanced: frequency-domain; best: deep learning model).
+**Tool route:** `noise-removal`  
+**Model:** SCUNet (tiered denoising pipeline)
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `quality` | `fast` \| `balanced` \| `best` | `balanced` | Processing tier |
-| `strength` | number (0–1) | 0.5 | Denoising strength |
-| `preserveDetail` | boolean | true | Edge-preserving mode |
-| `colorNoise` | boolean | false | Target color noise specifically |
+| `tier` | `"quick"` \| `"balanced"` \| `"quality"` \| `"maximum"` | `"balanced"` | Processing tier |
+| `strength` | number (0-100) | `50` | Denoising strength |
+| `detailPreservation` | number (0-100) | `50` | How much detail to preserve; higher keeps more texture |
+| `colorNoise` | number (0-100) | `30` | Color noise reduction strength |
+| `format` | string | `"original"` | Output format: `original`, `png`, `jpeg`, `webp`, `avif`, `jxl` |
+| `quality` | number (1-100) | `90` | Output encoding quality |
 
 ## Red Eye Removal
 
-**Function:** `removeRedEye`  
 **Tool route:** `red-eye-removal`
 
 Detects face landmarks, locates eye regions, and corrects red-channel oversaturation.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `sensitivity` | number (0–1) | 0.5 | Red pixel detection threshold |
-| `strength` | number (0–1) | 0.9 | Correction strength |
+| `sensitivity` | number (0-100) | `50` | Red pixel detection threshold |
+| `strength` | number (0-100) | `70` | Correction strength |
+| `format` | string | - | Output format override (optional) |
+| `quality` | number (1-100) | `90` | Output quality |
 
 ## Photo Restoration
 
-**Function:** `restorePhoto`  
 **Tool route:** `restore-photo`
 
-Multi-step pipeline for old or damaged photos: scratch/tear detection and repair → face enhancement → denoising → optional colorization.
+Multi-step pipeline for old or damaged photos: scratch/tear detection and repair, face enhancement, denoising, and optional colorization.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `mode` | `auto` \| `light` \| `heavy` | `auto` | Restoration intensity |
-| `scratchRemoval` | boolean | true | Detect and repair scratches, tears |
-| `faceEnhancement` | boolean | true | Apply face enhancement pass |
-| `fidelity` | number (0–1) | 0.7 | Face enhancement strength |
-| `denoise` | boolean | true | Apply denoising pass |
-| `denoiseStrength` | number (0–100) | 40 | Denoising strength |
-| `colorize` | boolean | false | Colorize after restoration |
+| `scratchRemoval` | boolean | `true` | Detect and repair scratches, tears |
+| `faceEnhancement` | boolean | `true` | Apply face enhancement pass |
+| `fidelity` | number (0-1) | `0.7` | Face enhancement strength (higher = more conservative) |
+| `denoise` | boolean | `true` | Apply denoising pass |
+| `denoiseStrength` | number (0-100) | `25` | Denoising strength |
+| `colorize` | boolean | `false` | Colorize after restoration |
+| `colorizeStrength` | number (0-100) | `85` | Colorization intensity |
 
 ## Passport Photo
 
-**Function:** Uses `detectFaceLandmarks` + `removeBackground`  
 **Tool route:** `passport-photo`  
-**Model:** MediaPipe face landmarks
+**Models:** MediaPipe face landmarks + BiRefNet background removal
 
-Generates government-compliant ID photos. Supports **37 countries** across 6 regions (Americas, Europe, Asia, Africa, Oceania, Middle East). Each spec includes physical dimensions, DPI, head-height ratio, eye-line position, and background color requirements.
+Two-phase workflow: analyze (detect face + remove background) then generate (crop, resize, tile). Supports 37+ countries across 6 regions.
+
+### Phase 1: Analyze
+
+`POST /api/v1/tools/passport-photo/analyze`
+
+Accepts an image file (multipart). Returns face landmark data, a base64 preview, and image dimensions.
+
+### Phase 2: Generate
+
+`POST /api/v1/tools/passport-photo/generate`
+
+Accepts a JSON body with the Phase 1 results plus generation settings:
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `country` | string | `us` | ISO country code (see list in UI) |
-| `printLayout` | `4x6` \| `A4` \| `none` | `none` | Output as print sheet or standalone |
-| `backgroundColor` | string | country default | Background fill color |
+| `jobId` | string | (required) | Job ID from Phase 1 |
+| `filename` | string | (required) | Original filename from Phase 1 |
+| `countryCode` | string | (required) | ISO country code (e.g., `US`, `GB`, `IN`) |
+| `documentType` | string | `"passport"` | Document type |
+| `bgColor` | string | `"#FFFFFF"` | Background color hex |
+| `printLayout` | string | `"none"` | Print layout: `none`, `4x6`, `a4`, `letter` |
+| `maxFileSizeKb` | number | `0` | Max file size in KB (0 = no limit) |
+| `dpi` | number (72-1200) | `300` | Output DPI |
+| `customWidthMm` | number | - | Custom width in mm (overrides country spec) |
+| `customHeightMm` | number | - | Custom height in mm (overrides country spec) |
+| `zoom` | number (0.5-3) | `1` | Zoom factor |
+| `adjustX` | number | `0` | Horizontal position adjustment |
+| `adjustY` | number | `0` | Vertical position adjustment |
+| `landmarks` | object | (required) | Landmarks from Phase 1 |
+| `imageWidth` | number | (required) | Image width from Phase 1 |
+| `imageHeight` | number | (required) | Image height from Phase 1 |
 
 ## Object Erasing (Inpainting)
 
-**Function:** `inpaint`  
 **Tool route:** `erase-object`  
 **Model:** LaMa via ONNX Runtime
 
-| Parameter | Type | Required | Description |
+The mask is sent as a **second file part** (fieldname `mask`), not as base64. White pixels in the mask indicate areas to erase. The `format` and `quality` settings are sent as top-level form fields.
+
+| Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `maskData` | string | Yes | Base64-encoded PNG mask (white = erase) |
-| `maskThreshold` | number (0–255) | No | Threshold for mask binarization |
+| `file` | file | (required) | Source image (multipart) |
+| `mask` | file | (required) | Mask image (multipart, fieldname `mask`, white = erase) |
+| `format` | string | `"auto"` | Output format: `auto`, `png`, `jpg`, `jpeg`, `webp`, `tiff`, `gif`, `avif`, `heic`, `heif`, `jxl` |
+| `quality` | integer (1-100) | `95` | Output quality |
 
 GPU-accelerated when an NVIDIA GPU is available.
 
-## Smart Crop
+## AI Canvas Expand
 
-**Function:** Uses MediaPipe + Sharp attention/entropy  
-**Tool route:** `smart-crop`  
-**Model:** MediaPipe face detection
+**Tool route:** `ai-canvas-expand`  
+**Model:** LaMa-based outpainting
+
+Expands the canvas of an image in any direction and fills new areas with AI-generated content that matches the existing image.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `mode` | `subject` \| `face` \| `trim` | `subject` | Crop strategy |
-| `width` | number | - | Output width |
-| `height` | number | - | Output height |
-| `facePreset` | string | - | Preset framing when `mode=face` |
+| `extendTop` | integer | `0` | Pixels to extend at the top |
+| `extendRight` | integer | `0` | Pixels to extend at the right |
+| `extendBottom` | integer | `0` | Pixels to extend at the bottom |
+| `extendLeft` | integer | `0` | Pixels to extend at the left |
+| `tier` | `"fast"` \| `"balanced"` \| `"high"` | `"balanced"` | Quality tier |
+| `format` | string | `"auto"` | Output format: `auto`, `png`, `jpg`, `jpeg`, `webp`, `tiff`, `gif`, `avif`, `heic`, `heif`, `jxl` |
+| `quality` | integer (1-100) | `95` | Output quality |
+
+At least one extend direction must be greater than 0.
+
+## Smart Crop
+
+**Tool route:** `smart-crop`  
+**Model:** MediaPipe face detection (face mode only)
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `mode` | string | `"subject"` | Crop strategy: `subject`, `face`, `trim` |
+| `strategy` | `"attention"` \| `"entropy"` | `"attention"` | Strategy for subject mode |
+| `width` | integer | - | Output width |
+| `height` | integer | - | Output height |
+| `padding` | integer (0-50) | `0` | Padding percentage around subject |
+| `facePreset` | string | `"head-shoulders"` | Preset framing when `mode=face` |
+| `sensitivity` | number (0-1) | `0.5` | Face detection threshold |
+| `threshold` | integer (0-255) | `30` | Background detection threshold (trim mode) |
+| `padToSquare` | boolean | `false` | Pad trimmed result to a square |
+| `padColor` | string | `"#ffffff"` | Background color for square padding |
+| `targetSize` | integer | - | Target size for padded output (pixels) |
+| `quality` | integer (1-100) | - | Output quality |
+
+Legacy `mode` values `attention` and `content` are accepted and mapped to `subject` and `trim` respectively.
 
 **Face presets:**
 
-| Preset | Head ratio | Best for |
-|--------|-----------|---------|
-| `close-up` | 1.8× face | Headshots |
-| `head-and-shoulders` | 2.8× face | Profile photos |
-| `upper-body` | 4.5× face | LinkedIn / formal |
-| `half-body` | 7.0× face | Full upper body |
+| Preset | Best for |
+|--------|---------|
+| `closeup` | Headshots |
+| `head-shoulders` | Profile photos |
+| `upper-body` | LinkedIn / formal |
+| `half-body` | Full upper body |
 
-## Image Enhancement
+## Transcribe Audio
 
-**Function:** `analyzeImage` + `applyCorrections`  
-**Tool route:** `image-enhancement`  
-**Engine:** Analysis-based (Sharp histogram and statistics)
+**Tool route:** `transcribe-audio`  
+**Model:** faster-whisper
 
-Analyzes the image and applies automatic corrections for exposure, contrast, white balance, saturation, sharpness, and noise. Supports scene-specific modes.
+Converts speech to text. Supports plain text, SRT, and VTT output formats.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `mode` | `auto` \| `portrait` \| `landscape` \| `low-light` \| `food` \| `document` | `auto` | Scene mode for tuning corrections |
-| `intensity` | number (0-100) | 50 | Overall correction strength |
-| `corrections.exposure` | boolean | true | Apply exposure correction |
-| `corrections.contrast` | boolean | true | Apply contrast correction |
-| `corrections.whiteBalance` | boolean | true | Apply white balance correction |
-| `corrections.saturation` | boolean | true | Apply saturation correction |
-| `corrections.sharpness` | boolean | true | Apply sharpness correction |
-| `corrections.denoise` | boolean | true | Apply denoising |
+| `language` | string | `"auto"` | Language: `auto`, `en`, `de`, `fr`, `es`, `zh`, `ja`, `ko`, `id`, `th`, `vi` |
+| `outputFormat` | `"txt"` \| `"srt"` \| `"vtt"` | `"txt"` | Output format |
 
-An additional analysis endpoint is available at `POST /api/v1/tools/image-enhancement/analyze` which returns the detected corrections without applying them.
+## Auto Subtitles
 
-## Content-Aware Resize (Seam Carving)
+**Tool route:** `auto-subtitles`  
+**Model:** faster-whisper (extracts audio from video, then transcribes)
 
-**Function:** `seamCarve`  
-**Tool route:** `content-aware-resize`  
-**Engine:** Go `caire` binary (not Python - no GPU benefit)
-
-Intelligently resizes images by removing or adding low-energy seams, preserving important content.
+Generates subtitle files from a video's audio track.
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `width` | number | - | Target width |
-| `height` | number | - | Target height |
-| `protectFaces` | boolean | true | Protect detected face regions from seam removal |
-| `blurRadius` | number | 0 | Pre-blur to reduce noise sensitivity |
-| `sobelThreshold` | number | 10 | Edge sensitivity threshold |
-| `square` | boolean | false | Force square output |
-
-Max input edge before auto-downscaling: **1200 px**.
+| `language` | string | `"auto"` | Language: `auto`, `en`, `de`, `fr`, `es`, `zh`, `ja`, `ko`, `id`, `th`, `vi` |
+| `format` | `"srt"` \| `"vtt"` | `"srt"` | Output subtitle format |
 
 ## PNG Transparency Fixer
 
-**Function:** `fixTransparency`  
 **Tool route:** `transparency-fixer`  
 **Model:** BiRefNet HR-matting (2048x2048 resolution)
 
@@ -271,12 +363,11 @@ Fixes "fake transparent" PNGs where the background was removed but left behind f
 
 **OOM fallback chain:** If BiRefNet HR-matting exceeds available memory, the tool automatically falls back to `birefnet-general`, then to `u2net`.
 
-**Feature bundle:** Background Removal (shared with Remove Background and Passport Photo).
-
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `defringe` | number (0-100) | 30 | Edge defringe strength to remove color contamination |
+| `defringe` | number (0-100) | `30` | Edge defringe strength to remove color contamination |
 | `outputFormat` | `"png"` \| `"webp"` | `"png"` | Output image format |
+| `removeWatermark` | boolean | `false` | Apply watermark removal pre-processing (median filter) |
 
 ```bash
 curl -X POST http://localhost:1349/api/v1/tools/transparency-fixer \
@@ -284,3 +375,46 @@ curl -X POST http://localhost:1349/api/v1/tools/transparency-fixer \
   -F "file=@fake-transparent.png" \
   -F 'settings={"defringe":30,"outputFormat":"png"}'
 ```
+
+---
+
+## Tools with Optional AI Capabilities
+
+The following tools are not Python sidecar tools but use AI features when certain options are enabled.
+
+### Image Enhancement
+
+**Tool route:** `image-enhancement`  
+**Engine:** Analysis-based (Sharp histogram and statistics)
+
+Analyzes the image and applies automatic corrections for exposure, contrast, white balance, saturation, sharpness, and noise. Supports scene-specific modes.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `mode` | `"auto"` \| `"portrait"` \| `"landscape"` \| `"low-light"` \| `"food"` \| `"document"` | `"auto"` | Scene mode for tuning corrections |
+| `intensity` | number (0-100) | `50` | Overall correction strength |
+| `corrections.exposure` | boolean | `true` | Apply exposure correction |
+| `corrections.contrast` | boolean | `true` | Apply contrast correction |
+| `corrections.whiteBalance` | boolean | `true` | Apply white balance correction |
+| `corrections.saturation` | boolean | `true` | Apply saturation correction |
+| `corrections.sharpness` | boolean | `true` | Apply sharpness correction |
+| `corrections.denoise` | boolean | `true` | Apply denoising |
+| `deepEnhance` | boolean | `false` | Enable AI noise removal via SCUNet (requires `upscale-enhance` bundle) |
+
+An additional analysis endpoint is available at `POST /api/v1/tools/image-enhancement/analyze` which returns the detected corrections without applying them.
+
+### Content-Aware Resize (Seam Carving)
+
+**Tool route:** `content-aware-resize`  
+**Engine:** Go `caire` binary (not Python -- no GPU benefit)
+
+Intelligently resizes images by removing low-energy seams, preserving important content.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `width` | number | - | Target width |
+| `height` | number | - | Target height |
+| `protectFaces` | boolean | `false` | Protect detected face regions (requires `face-detection` bundle) |
+| `blurRadius` | number (0-20) | `4` | Pre-blur for energy calculation |
+| `sobelThreshold` | number (1-20) | `2` | Edge sensitivity threshold |
+| `square` | boolean | `false` | Force square output |
diff --git a/apps/docs/api/image-engine.md b/apps/docs/api/image-engine.md
index dbcb6ef1..659b920f 100644
--- a/apps/docs/api/image-engine.md
+++ b/apps/docs/api/image-engine.md
@@ -32,6 +32,7 @@ Cut out a rectangular region from the image.
 | `top` | number | Y offset from the top edge |
 | `width` | number | Width of the crop area |
 | `height` | number | Height of the crop area |
+| `unit` | string | `px` (default) or `percent` |
 
 ### rotate
 
@@ -40,15 +41,16 @@ Rotate the image by a given angle.
 | Parameter | Type | Description |
 |---|---|---|
 | `angle` | number | Rotation angle in degrees (0-360) |
-| `background` | string | Fill color for the exposed area (default: transparent or white) |
+| `background` | string | Fill color for exposed area (default: `#000000`). Only applies to non-90-degree angles. |
 
 ### flip
 
-Mirror the image horizontally or vertically.
+Mirror the image horizontally, vertically, or both. At least one must be true.
 
 | Parameter | Type | Description |
 |---|---|---|
-| `direction` | string | `horizontal` or `vertical` |
+| `horizontal` | boolean | Mirror left to right |
+| `vertical` | boolean | Mirror top to bottom |
 
 ### convert
 
@@ -56,9 +58,11 @@ Change the image format.
 
 | Parameter | Type | Description |
 |---|---|---|
-| `format` | string | Target format: `jpeg`, `png`, `webp`, `avif`, `tiff`, `gif`, `jxl`, `heic`, `heif`, `bmp`, `ico`, `jp2`, `qoi` |
+| `format` | string | Target format: `jpg`, `png`, `webp`, `avif`, `tiff`, `gif`, `jxl`, `heic`, `heif`, `bmp`, `ico`, `jp2`, `qoi` |
 | `quality` | number | Compression quality (1-100, applies to lossy formats) |
 
+The first seven formats (`jpg` through `jxl`) are encoded by Sharp in-process. The remaining formats use external encoders at the API layer: `heic`/`heif` via heif-enc, `bmp`/`ico` via ImageMagick, `jp2` via opj_compress, and `qoi` via an inline TypeScript codec.
+
 ### compress
 
 Reduce file size while keeping the same format.
@@ -66,11 +70,20 @@ Reduce file size while keeping the same format.
 | Parameter | Type | Description |
 |---|---|---|
 | `quality` | number | Target quality (1-100) |
+| `targetSizeBytes` | number | Optional target file size in bytes |
 | `format` | string | Optional format override |
 
 ### strip-metadata
 
-Remove EXIF, IPTC, and XMP metadata from the image. Useful for privacy before sharing photos publicly. Takes no parameters.
+Remove EXIF, IPTC, XMP, and ICC metadata from the image. With no parameters (or `stripAll: true`), strips everything. Pass individual flags for selective stripping.
+
+| Parameter | Type | Description |
+|---|---|---|
+| `stripAll` | boolean | Strip all metadata (default when no flags are set) |
+| `stripExif` | boolean | Strip EXIF data (including GPS if `stripGps` is not separately set) |
+| `stripGps` | boolean | Strip GPS location data |
+| `stripIcc` | boolean | Strip ICC color profile |
+| `stripXmp` | boolean | Strip XMP metadata |
 
 ### Color adjustments
 
@@ -94,35 +107,102 @@ These apply a fixed color transformation. They take no parameters.
 
 ### Color channels
 
-Adjust individual RGB color channels.
+Adjust individual RGB color channels. Values are multipliers where 100 = no change.
 
 | Parameter | Type | Description |
 |---|---|---|
-| `red` | number | Red channel adjustment (-100 to 100) |
-| `green` | number | Green channel adjustment (-100 to 100) |
-| `blue` | number | Blue channel adjustment (-100 to 100) |
+| `red` | number | Red channel multiplier (0 to 200, 100 = unchanged) |
+| `green` | number | Green channel multiplier (0 to 200, 100 = unchanged) |
+| `blue` | number | Blue channel multiplier (0 to 200, 100 = unchanged) |
+
+### sharpen
+
+Simple sharpening controlled by a single value.
+
+| Parameter | Type | Description |
+|---|---|---|
+| `value` | number | Sharpening intensity (0 to 100). Mapped to a Gaussian sigma of 0.5-10. |
+
+### sharpen-advanced
+
+Advanced sharpening with three selectable methods and an optional noise-reduction pre-pass.
+
+| Parameter | Type | Description |
+|---|---|---|
+| `method` | string | `adaptive`, `unsharp-mask`, or `high-pass` |
+| `sigma` | number | Gaussian blur radius, 0.5-10 (adaptive) |
+| `m1` | number | Flat-area sharpening, 0-10 (adaptive) |
+| `m2` | number | Textured-area sharpening, 0-20 (adaptive) |
+| `x1` | number | Flat/jagged threshold, 0-10 (adaptive) |
+| `y2` | number | Max brightening (halo clamp), 0-50 (adaptive) |
+| `y3` | number | Max darkening (halo clamp), 0-50 (adaptive) |
+| `amount` | number | Intensity percentage, 0-500 (unsharp-mask) |
+| `radius` | number | Blur radius, 0.1-5.0 (unsharp-mask) |
+| `threshold` | number | Minimum edge brightness, 0-255 (unsharp-mask) |
+| `strength` | number | Blend strength, 0-100 (high-pass) |
+| `kernelSize` | number | `3` or `5` for 3x3 / 5x5 kernel (high-pass) |
+| `denoise` | string | Noise reduction pre-pass: `off`, `light`, `medium`, or `strong` |
+
+Parameters are method-specific. Only supply the ones relevant to the chosen method.
+
+### color-blindness
+
+Simulate a color vision deficiency using a 3x3 color-recombination matrix.
+
+| Parameter | Type | Description |
+|---|---|---|
+| `type` | string | One of: `protanopia`, `deuteranopia`, `tritanopia`, `protanomaly`, `deuteranomaly`, `tritanomaly`, `achromatopsia`, `blueConeMonochromacy` |
+
+### edit-metadata
+
+Write or remove individual EXIF/IPTC metadata fields without stripping the entire block.
+
+| Parameter | Type | Description |
+|---|---|---|
+| `artist` | string | EXIF Artist tag |
+| `copyright` | string | EXIF Copyright tag |
+| `imageDescription` | string | EXIF ImageDescription tag |
+| `software` | string | EXIF Software tag |
+| `dateTime` | string | EXIF DateTime tag |
+| `dateTimeOriginal` | string | EXIF DateTimeOriginal tag |
+| `clearGps` | boolean | Remove all GPS tags |
+| `fieldsToRemove` | string[] | List of EXIF field names to delete |
+
+All parameters are optional. Fields listed in `fieldsToRemove` are deleted from the existing EXIF block. Fields set via the named parameters are written (or overwritten). Binary/unsafe keys like MakerNote are silently ignored.
 
 ## Format detection
 
 The engine detects input formats automatically from file headers, not just file extensions. This means a `.jpg` file that is actually a PNG will be handled correctly. Detection uses a multi-layer approach: magic bytes first, then file extension as fallback.
 
-SnapOtter supports **55+ input formats** and **14 output formats**, including 23 camera RAW formats from 20+ brands, professional formats (PSD, EPS, OpenEXR, HDR), modern codecs (JPEG XL, AVIF, HEIC, QOI, JPEG 2000), and scientific/gaming formats (FITS, DDS). Decoding is handled by Sharp natively where possible, with automatic fallback to ImageMagick, LibRaw, and specialized CLI decoders.
+SnapOtter supports **55+ input formats** and **13 output formats**, including 23 camera RAW formats from 20+ brands, professional formats (PSD, EPS, OpenEXR, HDR), modern codecs (JPEG XL, AVIF, HEIC, QOI, JPEG 2000), and scientific/gaming formats (FITS, DDS). Decoding is handled by Sharp natively where possible, with automatic fallback to ImageMagick, LibRaw, and specialized CLI decoders.
 
 See the [Supported Formats](/guide/supported-formats) page for the complete list.
 
 ## Metadata extraction
 
-The `info` tool returns image metadata:
+The `info` tool returns image metadata. See [Image Info](/tools/image/info) for the full field reference.
 
 ```json
 {
-  "width": 1920,
-  "height": 1080,
+  "filename": "photo.jpg",
+  "fileSize": 2450000,
+  "width": 4032,
+  "height": 3024,
   "format": "jpeg",
-  "size": 245678,
   "channels": 3,
   "hasAlpha": false,
-  "dpi": 72,
-  "exif": { ... }
+  "colorSpace": "srgb",
+  "density": 72,
+  "isProgressive": false,
+  "hasExif": true,
+  "hasIcc": true,
+  "hasXmp": false,
+  "bitDepth": "8",
+  "pages": 1,
+  "histogram": [
+    { "channel": "red", "min": 0, "max": 255, "mean": 128.45, "stdev": 52.31 },
+    { "channel": "green", "min": 2, "max": 253, "mean": 115.22, "stdev": 48.76 },
+    { "channel": "blue", "min": 0, "max": 250, "mean": 102.89, "stdev": 55.14 }
+  ]
 }
 ```
diff --git a/apps/docs/api/rest.md b/apps/docs/api/rest.md
index a18ccbe5..a3d272f4 100644
--- a/apps/docs/api/rest.md
+++ b/apps/docs/api/rest.md
@@ -102,8 +102,9 @@ curl -X POST http://localhost:1349/api/v1/tools/<toolId>/batch \
 
 - Upload is `multipart/form-data`.
 - `settings` is a JSON string with tool-specific options.
-- Response is the processed file directly (or a ZIP for batch).
-- Progress is tracked via SSE (see [Progress Tracking](#progress-tracking)).
+- **Fast tools** (200) return JSON: `{"jobId":"...","downloadUrl":"/api/v1/download/<jobId>/<filename>","originalSize":1234,"processedSize":567}`. Fetch the processed file from `downloadUrl`.
+- **Long-running tools** (202) return JSON: `{"jobId":"...","async":true}`. Connect to SSE for progress, then download when complete (see [Progress Tracking](#progress-tracking)).
+- **Batch** returns a ZIP archive streamed directly (with `X-Job-Id` header).
 
 ## Tools Reference
 
@@ -163,6 +164,7 @@ All AI tools run on your hardware (CPU or NVIDIA GPU). No internet required.
 | `transparency-fixer` | PNG Transparency Fixer | BiRefNet HR-matting | `defringe` (0-100), `outputFormat` (png/webp) |
 | `background-replace` | Background Replace | rembg (BiRefNet) | `backgroundType` (color/gradient), `color` (hex), `gradientColor1`, `gradientColor2`, `gradientAngle`, `feather` (0-20), `format` (png/webp) |
 | `blur-background` | Blur Background | rembg (BiRefNet) | `intensity` (1-100), `feather` (0-20), `format` (png/webp) |
+| `ai-canvas-expand` | AI Canvas Expand | LaMa (outpainting) | `extendTop`, `extendRight`, `extendBottom`, `extendLeft` (px), `tier` (fast/balanced/high), `format`, `quality` |
 
 ### Watermark & Overlay
 
@@ -435,7 +437,7 @@ Each step's output is the next step's input. Unlimited steps per pipeline by def
 Long-running jobs (AI tools, batch, pipelines) emit real-time progress via Server-Sent Events:
 
 ```bash
-# Connect to the SSE stream (jobId returned in X-Job-Id response header)
+# Connect to the SSE stream (jobId is in the JSON response body from the tool endpoint)
 curl -N http://localhost:1349/api/v1/jobs/<jobId>/progress \
   -H "Authorization: Bearer <token>"
 ```
@@ -505,7 +507,7 @@ Custom role management with granular permissions.
 | `PUT` | `/api/v1/roles/:id` | Admin (`users:manage`) | Update a custom role (cannot modify built-in roles) |
 | `DELETE` | `/api/v1/roles/:id` | Admin (`users:manage`) | Delete a custom role (cannot delete built-in roles; affected users revert to `user` role) |
 
-Available permissions: `tools:use`, `files:own`, `files:all`, `apikeys:own`, `apikeys:all`, `pipelines:own`, `pipelines:all`, `settings:read`, `settings:write`, `users:manage`, `teams:manage`, `features:manage`, `system:health`, `audit:read`.
+Available permissions (17): `tools:use`, `files:own`, `files:all`, `apikeys:own`, `apikeys:all`, `pipelines:own`, `pipelines:all`, `settings:read`, `settings:write`, `users:manage`, `teams:manage`, `features:manage`, `system:health`, `audit:read`, `compliance:manage`, `webhooks:manage`, `security:manage`.
 
 ## Audit Log
 
diff --git a/apps/docs/guide/supported-formats.md b/apps/docs/guide/supported-formats.md
index c17191b1..5dfeee04 100644
--- a/apps/docs/guide/supported-formats.md
+++ b/apps/docs/guide/supported-formats.md
@@ -8,7 +8,7 @@ SnapOtter processes files across five modalities: image, video, audio, document,
 
 ## Image Formats
 
-SnapOtter supports 55+ image formats for input and 14 formats for output.
+SnapOtter supports 55+ image formats for input and 13 formats for output.
 
 ## Input Formats
 
@@ -101,7 +101,7 @@ SnapOtter supports 55+ image formats for input and 14 formats for output.
 | PAM | .pam | Sharp (native) | Arbitrary map |
 | PFM | .pfm | Sharp (native) | Float map |
 
-## Output Formats (14)
+## Output Formats (13)
 
 | Format | Encoder | Quality Control | Available In |
 |--------|---------|----------------|-------------|