Migrate CPU-intensive workloads to Cloud Run for autoscaling:
- Add Whisper HTTP service (FastAPI) with /transcribe endpoint
- Add FFmpeg HTTP service (FastAPI) with /encode, /probe, /extract-frame, etc.
- Add Dockerfiles for both services (8 vCPU, 32GB RAM, Gen2)
- Add Cloud Build config for CI/CD deployment
- Add Cloud Run service YAML configs with scale-to-zero
- Update whisper_transcribe.py to call Cloud Run when WHISPER_SERVICE_URL set
- Update video_renderer.py to call Cloud Run when FFMPEG_SERVICE_URL set
- Update whisper_service.py for Cloud Run compatibility (no settings dependency)
- Add ffmpeg_service_url and whisper_service_url to config.py
Services scale 0→N based on request load, falling back to local
execution when service URLs are not configured (hybrid mode).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Use explicit null returns and String() casts for unknown types
from job.error Record to satisfy TypeScript strict mode.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add RENDER_FAILED job status for when video rendering fails
- Fix _check_accessible_video_completion to detect failures and transition
job status accordingly (was stuck in RENDERING_VIDEO forever)
- Store detailed error info in job.error including failed_languages array
- Call completion check after failures to properly update job status
- Broadcast WebSocket notification on render failures
Frontend:
- Add render_failed to JobStatus type and StatusBadge (red styling)
- Add tts_failed and render_failed to JobsList STATUS_LABELS
- Enhance JobDetail error display with:
- Warning icon and prominent styling
- Error type and message
- Failed languages list with per-language errors
- Timestamp of when error occurred
- Update ProgressIndicator to handle failed states with red dot
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changed "Target Languages" to "Target Languages for Translation"
for clarity in the new job form.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Languages column now shows total languages (source + translations)
instead of just translation count, matching other parts of the UI.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Video-native translation mode now processes all target languages in parallel
using asyncio.gather() with a semaphore (max 3 concurrent) for rate limiting.
This significantly reduces total translation time when multiple languages
are selected.
- Add MAX_CONCURRENT_VIDEO_NATIVE constant for rate limiting
- Refactor video-native path to use parallel coroutines
- Keep traditional VTT translation mode sequential
- Handle per-language errors without stopping other translations
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The google-cloud-storage library made blob.generation read-only,
causing job deletion to fail silently (0 GCS files deleted).
Using bucket.delete_blob(name) instead avoids generation checking.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add a new "Video Native Mode" translation option that re-processes the
video through Gemini for each target language, generating captions and
audio descriptions directly from visual context. This produces more
natural and culturally appropriate content compared to traditional VTT
text translation.
Changes:
- Add translation_mode field to RequestedOutputs (video_native | traditional)
- Create gemini_ingestion_targeted.md prompt for target language generation
- Add extract_accessibility_targeted() method to Gemini service
- Modify translate_and_synthesize task to handle both translation modes
- Add Translation Mode UI selector in NewJob screen (video_native is default)
- Remove transcreation UI (replaced by video_native mode)
- Remove Google Translate service (replaced by Gemini translation)
- Add LanguageSelector component with searchable dropdown
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Deduplicate job IDs to prevent processing same job twice
- Convert GCS blob iterator to list upfront to avoid stale generations
- Clear blob.generation before delete to handle concurrent deletions
- Catch NotFound errors gracefully for already-deleted blobs
- Don't re-raise GCS errors - cleanup failures shouldn't block deletion
- Treat already-deleted jobs as successful (idempotent delete)
- Disable action dropdown during bulk operations in UI
- Show spinner with "Please wait" message during deletion
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add comprehensive error handling for TTS synthesis failures:
Backend:
- Add TTS_FAILED status to JobStatus enum for failed synthesis jobs
- Add TTSSynthesisError exception with cue index and context tracking
- Improve null-safe error handling in Gemini TTS response parsing
- Add _synthesize_cue_with_retry() with exponential backoff (3 attempts)
- Enhanced error logging with text preview and model context
Frontend:
- Add TTS_FAILED status styling (red badge) in StatusBadge component
- Add tts_failed to JobStatus TypeScript type
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Break out TTS synthesis into a dedicated Celery worker (tts queue) with
concurrency=8 for parallel processing. Each AD cue is now synthesized as
a separate task, enabling up to 8 cues to be processed simultaneously.
Key changes:
- Add tts_synthesis.py with synthesize_cue_task for per-cue synthesis
- Refactor translate_and_synthesize.py to dispatch cue tasks in parallel
- Add tts-worker service to docker-compose.yml (concurrency=8)
- Add Cloud Run service config for production deployment
Benefits:
- Parallel synthesis even for single jobs (e.g., 50 cues → 8 concurrent)
- Natural rate limiting across multiple concurrent jobs
- Fault tolerance with per-cue retries and GCS persistence
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The VTT retimer had two bugs causing subtitles to display during freeze
periods and become out of sync:
1. Same offset applied to both start and end times (should differ when
pause falls between them)
2. Cues spanning pause points weren't split (causing captions during freeze)
Changes:
- Add _offset_at() for timestamps AT or AFTER pause points
- Add _offset_before() for timestamps STRICTLY BEFORE pause points
- Add _retime_cue() to split cues at pause points into multiple segments
- Add _filter_short_segments() to remove <100ms segments after splitting
- Rewrite retime_for_pause_insert() to use new helper methods
Example fix for cue 8s-12s with pause at 10s (4s freeze):
- Before: 8s-12s (displayed during freeze!)
- After: 8s-10s + 14s-16s (gap during AD)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The video event listener effect had an empty dependency array, causing it
to run only once on mount. For batch uploads where downloads load slower,
the video element wasn't rendered yet when the effect ran, so listeners
were never attached. Adding currentVideoUrl as a dependency ensures the
effect re-runs when the video becomes available.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace complex overlap/catch-up logic with simpler approach:
- Snap pause points to midpoint between sentences (not sentence boundaries)
- Add 500ms silence before AND after AD audio during freeze frame
- Resume playback from same midpoint (no overlap, no visual jump-back)
This eliminates audio/visual anomalies caused by the previous algorithm's
complexity around sentence boundary snapping and audio catch-up.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Get video duration BEFORE the render loop (not after)
- Clamp pause_point to 100ms before video end if it exceeds duration
- Add validation in _extract_frame() to verify frame was created
- Add debug logging for frame extraction timestamps
This prevents "Frame file not found" errors when pause points
calculated by Whisper refinement exceed the source video duration.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When a pause point falls between two sentences, the previous algorithm
created a visual jump-back where the video rewound to resume_from after
the AD played. This was distracting to viewers.
New behavior:
- Video plays normally to pause_point
- Freeze frame shows + AD audio plays
- Freeze frame CONTINUES while source audio from [resume_from, pause_point]
plays (the "catch-up" audio)
- Video resumes smoothly from pause_point (no visual jump)
The audio from the overlap region plays twice (once during video, once
during freeze extension) but this is acceptable as it's typically <1s
and provides natural audio context around the AD.
Implementation:
- Add _extract_audio_segment() to extract catch-up audio from source
- Add _concatenate_audio() to join AD + catch-up audio
- Modify render loop to create extended freeze segments with combined audio
- Resume video from pause_point instead of resume_from
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Show last 1500 chars of stderr instead of first 500 to capture actual
error messages (FFmpeg writes version banner first, errors at end)
- Add validation for freeze segment creation:
- Check duration > 0
- Verify frame and audio files exist
- Add debug logging for parameters
This helps diagnose FFmpeg failures that were previously showing only
version/configuration info instead of the actual error.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Completely rewrites the Whisper-based pause point refinement to use
a two-phase approach with explicit ordering:
Phase 1 - Individual refinement:
1. Check if pause point is "during speaking" (words within ±2s)
- If NOT during speaking → use Gemini's exact point, no overlap
2. If during speaking, find nearest sentence boundary
3. Apply appropriate buffering based on context:
- Case A: First sentence → pause 500ms before sentence starts
- Case B: Last sentence → pause 500ms after sentence ends
- Case C: Between sentences → full double buffer (overlap)
Phase 2 - Consolidation (after all refinements):
- Consolidate cues within 5s of each other to play back-to-back
Key changes:
- Add SentenceBoundary dataclass for tracking boundaries with context
- Add _is_during_speaking() helper to detect speech proximity
- Add _find_sentence_boundaries() with longest-gap fallback
- Rewrite snap_pause_point() with new ordered algorithm
- Update refine_all_pause_points() to pass words and use two phases
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously, all consolidated cues shared the same pause_point AND
resume_from, which caused the overlap video segment to play between
each AD cue in a consolidated group.
Now consolidated cues are treated as a single AD segment:
- All cues in a group share the same pause_point (front buffer once)
- Only the LAST cue keeps resume_from (back buffer once)
- Other cues have resume_from = pause_point (no video between ADs)
This ensures consolidated ADs play seamlessly back-to-back:
- Video plays up to pause_point (front buffer)
- AD_1 plays
- AD_2 plays immediately (no video)
- AD_n plays immediately (no video)
- Video resumes from resume_from (back buffer)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Remove the cached transcript lookup - always run a fresh Whisper
transcription for each accessible video render. This ensures we get
accurate word timestamps for the current video file.
The transcript is still saved to the job document for debugging and
auditing purposes, but it will never be read back for reuse.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When the Whisper analysis detects no speech near a Gemini-recommended
pause point, skip the full-gap-overlap algorithm and use the exact
pause point with no overlap (pause_point == resume_from).
This handles cases where Gemini chose a pause point in a silent or
music-only section of the video - there's no dialogue to buffer
around, so we simply pause and resume at the exact same point.
Three outcomes now in snap_pause_point():
1. No speech nearby → exact pause point, no overlap, no warning
2. Speech but no sentence break → warning (existing behavior)
3. Sentence break found → full-gap-overlap (existing behavior)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changes pause point calculation to use the entire gap between sentences
as a buffer on BOTH sides of the audio description:
- pause_point: Just BEFORE next sentence starts (gap.end - 50ms)
- resume_from: Just AFTER previous sentence ends (gap.start + 50ms)
This means a small portion of video plays twice (the gap duration), but
creates a much more natural listening experience by maximizing the
breathing room around audio descriptions.
Changes:
- whisper_service.py: snap_pause_point() now returns (pause_point, resume_from)
- video_renderer.py: Uses resume_from for current_time after freeze segment
- vtt_retimer.py: Calculates effective_offset including overlap duration
- accessible_video.py: Added resume_from field to ADPlacementCue schema
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Logs pause point placements, segment creation, and final segment
calculation to help diagnose the 30s black footage issue.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Medium model is faster and uses less memory (~1.5GB vs ~3GB)
while still providing good multilingual transcription quality.
Updated in:
- config.py
- docker-compose.yml
- whisper-worker-service.yaml
- cloudbuild.yaml
- Dockerfile (pre-download)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Instead of a fixed 175ms buffer, the pause point is now placed
halfway between the end of the sentence and the start of the
next word. If the half-gap exceeds 2 seconds, uses 500ms instead.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Logs VTT content fetching, parsing, and current caption state
to help diagnose subtitle overlay issues in final review mode.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Dynamically detects CPU count with os.cpu_count() instead of
hardcoded 4 threads. Falls back to 4 if detection fails.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Log model name explicitly when loading and transcribing
- Log model load time
- Log transcription processing time
- Helps verify correct model is being used
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Environment variables were overriding config.py with 'base' model.
Updated defaults in:
- docker-compose.yml
- whisper-worker-service.yaml
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The setting in config.py (5.0) was overriding the default in
whisper_service.py (30.0). Now both are consistent at 30s.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Downloads the model (~3GB) at build time to avoid cold start delays.
Also updated comment to reflect large-v3 memory usage (~4-6GB).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Uses the multilingual large model for more accurate transcription
and sentence boundary detection.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
If consecutive AD cues have pause points within 5 seconds, they now
play back-to-back at the same pause point. This prevents AD from being
inserted mid-sentence when cues are close together.
Adds _consolidate_close_cues() method and consolidation_threshold
parameter to refine_all_pause_points().
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Increases the search window from ±10s to ±20s to maximize the chance
of finding a proper sentence ending and avoid falling back to Gemini's
potentially imprecise pause points.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When the first AD cue (index 0) cannot find a sentence boundary within
the ±10s search window, insert the AD at T=0:00 instead of using the
potentially mid-sentence Gemini pause point.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update Gemini prompt to require transcription with precise timestamps
- Add sentence_boundaries output field for validation
- Add pause_point_rationale field to explain each pause point choice
- Emphasize terminal punctuation only (., ?, !) - never commas
- Expand Whisper search window from ±5s to ±10s
- Increase post-pause buffer from 50ms to 175ms
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The "Download All Files" function was missing accessible_video_mp4 and
accessible_captions_vtt files that the backend provides. Updated both
the bulk download in JobsList and the individual Downloads page to
include all available file types.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace 3-stage redundant deletion with single prefix-based approach.
All job files are under {job_id}/ prefix, so listing and deleting by
prefix is simpler and catches all files including new types like
accessible_video.mp4 and ad_cues/*.mp3.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Updated pause point algorithm:
- Search range: 5 seconds BEFORE to 5 seconds AFTER Gemini pause point
- ONLY considers sentence breaks (after periods, !, ?) - not phrase breaks
- Chooses the closest sentence break to the Gemini pause point
This ensures audio descriptions are inserted at natural sentence
boundaries, not in the middle of sentences after commas.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Two fixes:
1. Snap pause point to gap.start (end of previous word) to prevent
cutting off the first word after the pause
2. Add explicit whisper_transcribe import in celery_worker.py
3. Fix misleading queue log message
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The pause point algorithm was snapping to gap.end (start of next word),
which caused the first word after the pause to be cut off. Changed to
snap to gap.start (end of previous word) instead.
Now the video pauses right after a word finishes, the AD plays during
the silence gap, and the next word plays in full when video resumes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Changed the Whisper transcription to run on dedicated whisper-worker
using the same dispatch pattern as FFmpeg:
1. apply_async() to dispatch to the whisper queue
2. Poll with ready() using async sleep to avoid blocking
3. Use allow_join_result() context manager
4. Get result only after task is ready
This ensures Whisper runs with concurrency=1 on a dedicated worker
to prevent memory overload while still allowing the render task
to wait for results without deadlocking.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Celery does not allow calling result.get() within a task as it causes
deadlocks. Changed the implementation to run Whisper transcription
directly using asyncio.to_thread() instead of dispatching to a separate
Celery queue.
The Whisper transcript is still cached in MongoDB for reuse across
language variants.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The rendering_video status was defined in job.py and frontend types but
was missing from the MongoDB schema validator, causing document update
failures when jobs transitioned to the rendering_video state.
Changes:
- Add migration script to update existing databases
- Update mongodb-init.js for new database setups
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
