03 / 06 Backend Developer live demo below

AI-Enhanced Multilingual Press Release Generator

Crafted an AI-powered tool using Flask and Python to transform PIB press releases into engaging multilingual videos, increasing accessibility. Established a centralized image management system and advanced prompt engineering to streamline video creation workflows, cutting production time by 45% and boosting productivity by 35%.

Source on GitHub ↗Jump to the demo ↓

01 · The constraint
A billion-person audience, a Hindi-only output pipe
The Press Information Bureau publishes 200-odd statements a day across 50+ ministries. The English text reaches the press; the actual audience — including the ~70 % of India that doesn't read English comfortably — is served by manual translators producing a trickle of videos for one or two regional languages.
The throughput gap was the bottleneck this project attacks: not the writing, the delivery of writing as a watchable, listenable artefact in every official language the citizen actually speaks.
02 · Ingestion
A Selenium walker pulls the daily docket
scrape_pib.py opens pib.gov.in/allRel.aspx headless, picks today's date, and walks the DOM: ministries → press-release lists → detail pages. Each detail page yields a title, a body, and a list of inline image URLs.
The output is an OriginalPressRelease per item, written into MongoDB keyed by the upstream PRID. Re-running the scraper is idempotent — PRIDs already in Mongo are skipped before any AI cost is spent.
03 · Summarise
GPT turns paragraphs into slide-sized chunks (and image keywords)
Raw press releases are long, dense, and not built for a 30-second video. DescriptiveContentGenerator prompts GPT for two things: a list of paragraph-per-slide summaries (≤2000 total chars) and a list of image-search keywords pulled from the news.
The keyword list is the bridge to step 4 — scrape_images.py queries Bing for each keyword and merges the results into the same DescriptiveContent record. Slides + relevant images, ready for compose.
●Models/DescriptiveContent.pypython
```
1prompt = f"""Given the press release: "{pr.content}",
2generate more descriptive and summarized content...
3 
4Result fields:
5  descriptive_text: list[str]  # one paragraph per slide
6  key_words:        list[str]  # 5-10 image-search terms
7  language:         "english"
8The sum of all descriptive_text chars must not exceed 2000.
9Output JSON only."""
10response = MetaAI().prompt(prompt)
```
04 · Multilingual fan-out
10 languages in parallel · googletrans + gTTS
convert_to_speech_in_multiple_languages(prId, joined_text, languages) runs googletrans for each target language (en, hi, bn, te, mr, ta, ur, gu, ml, kn), then gTTS to render an MP3 per language. The 10 audio files land on disk keyed by {prId}_output_{lang}.mp3.
Same upstream the browser demo uses — /api/tts proxies Google's translate_tts endpoint to mirror gTTS's behaviour, and /api/translate proxies MyMemory because Google Translate's free API has a stricter quota.
●text_to_voice.pypython
```
1def convert_to_speech_in_multiple_languages(prId, phrase, languages):
2    translator = Translator()
3    for lang in languages:
4        try:
5            translated = translator.translate(phrase, dest=lang).text
6            tts = gTTS(text=translated, lang=lang)
7            tts.save(f"{prId}_output_{lang}.mp3")
8        except Exception as e:
9            print(f"Error :{e} for language:{lang}")
```
05 · Compose
MoviePy: zoom + blur + fade · title overlay at y=0.05
generate_pr_video.GeneratePRVideo is the heavy hitter. For each image: a blurred background clip, a foreground clip scaled to 1.0 → 1.2× over the slide via the custom GradualScaleClip, 1 s fadein + fadeout. The translated title sits on a semi-transparent dark box at y_pct=0.05. Slide duration = audio length ÷ image count.
Final clip is concatenated, audio attached, written out as {prId}.mp4 at 1920×1080 / 30 fps via libx264. The browser demo paints the same effects (zoom, fade, title overlay) onto Canvas frame-by-frame instead of MoviePy.
●generate_pr_video.pypython
```
1image_clip = VideoFileClip(image_file, audio=False)
2image_clip = image_clip.resize(scaling_factor)
3scaled_clip = GradualScaleClip(
4    image_clip, zoom_factor=1.2,
5    duration=duration_per_image, target_fps=video_fps)
6fade_in_clip  = fadein(scaled_clip, duration=1)
7fade_out_clip = fadeout(scaled_clip, duration=1)
8text_line = TextClip(translator.translate(pr.title, dest=lan).text,
9                     font="Segoe-UI", fontsize=40, color="white")
```
06 · Distribute
Firebase Storage + FCM · 45 % less production time
Output MP4 and the ten MP3s go to Firebase Storage under gs://synth-ai-envoys.appspot.com/. Each new DescriptiveContent triggers an FCM push to every user subscribed to that ministry — title: "New Press Release", body and thumbnail derived from the record.
Net effect on the production team: ~30 videos a day, ten languages each, with the human only intervening for moderation. The 45 % time reduction and 35 % throughput lift in the project description come from this stage replacing all of the previous manual work.
07 · Try it
Run the six stages in your browser
Paste a press-release-style paragraph below. The pipeline runs the same shape as the Flask service — ingest, summarise (paragraph-split here, LLM in prod), translate to all 10 languages in parallel via /api/translate, speak the chosen language via /api/tts (gTTS upstream), compose a Canvas slideshow with the same zoom + fade + title overlay MoviePy produced, then a simulated upload.

01 · The constraint
A billion-person audience, a Hindi-only output pipe
The Press Information Bureau publishes 200-odd statements a day across 50+ ministries. The English text reaches the press; the actual audience — including the ~70 % of India that doesn't read English comfortably — is served by manual translators producing a trickle of videos for one or two regional languages.
The throughput gap was the bottleneck this project attacks: not the writing, the delivery of writing as a watchable, listenable artefact in every official language the citizen actually speaks.
02 · Ingestion
A Selenium walker pulls the daily docket
scrape_pib.py opens pib.gov.in/allRel.aspx headless, picks today's date, and walks the DOM: ministries → press-release lists → detail pages. Each detail page yields a title, a body, and a list of inline image URLs.
The output is an OriginalPressRelease per item, written into MongoDB keyed by the upstream PRID. Re-running the scraper is idempotent — PRIDs already in Mongo are skipped before any AI cost is spent.
03 · Summarise
GPT turns paragraphs into slide-sized chunks (and image keywords)
Raw press releases are long, dense, and not built for a 30-second video. DescriptiveContentGenerator prompts GPT for two things: a list of paragraph-per-slide summaries (≤2000 total chars) and a list of image-search keywords pulled from the news.
The keyword list is the bridge to step 4 — scrape_images.py queries Bing for each keyword and merges the results into the same DescriptiveContent record. Slides + relevant images, ready for compose.
●Models/DescriptiveContent.pypython
```
1prompt = f"""Given the press release: "{pr.content}",
2generate more descriptive and summarized content...
3 
4Result fields:
5  descriptive_text: list[str]  # one paragraph per slide
6  key_words:        list[str]  # 5-10 image-search terms
7  language:         "english"
8The sum of all descriptive_text chars must not exceed 2000.
9Output JSON only."""
10response = MetaAI().prompt(prompt)
```
04 · Multilingual fan-out
10 languages in parallel · googletrans + gTTS
convert_to_speech_in_multiple_languages(prId, joined_text, languages) runs googletrans for each target language (en, hi, bn, te, mr, ta, ur, gu, ml, kn), then gTTS to render an MP3 per language. The 10 audio files land on disk keyed by {prId}_output_{lang}.mp3.
Same upstream the browser demo uses — /api/tts proxies Google's translate_tts endpoint to mirror gTTS's behaviour, and /api/translate proxies MyMemory because Google Translate's free API has a stricter quota.
●text_to_voice.pypython
```
1def convert_to_speech_in_multiple_languages(prId, phrase, languages):
2    translator = Translator()
3    for lang in languages:
4        try:
5            translated = translator.translate(phrase, dest=lang).text
6            tts = gTTS(text=translated, lang=lang)
7            tts.save(f"{prId}_output_{lang}.mp3")
8        except Exception as e:
9            print(f"Error :{e} for language:{lang}")
```
05 · Compose
MoviePy: zoom + blur + fade · title overlay at y=0.05
generate_pr_video.GeneratePRVideo is the heavy hitter. For each image: a blurred background clip, a foreground clip scaled to 1.0 → 1.2× over the slide via the custom GradualScaleClip, 1 s fadein + fadeout. The translated title sits on a semi-transparent dark box at y_pct=0.05. Slide duration = audio length ÷ image count.
Final clip is concatenated, audio attached, written out as {prId}.mp4 at 1920×1080 / 30 fps via libx264. The browser demo paints the same effects (zoom, fade, title overlay) onto Canvas frame-by-frame instead of MoviePy.
●generate_pr_video.pypython
```
1image_clip = VideoFileClip(image_file, audio=False)
2image_clip = image_clip.resize(scaling_factor)
3scaled_clip = GradualScaleClip(
4    image_clip, zoom_factor=1.2,
5    duration=duration_per_image, target_fps=video_fps)
6fade_in_clip  = fadein(scaled_clip, duration=1)
7fade_out_clip = fadeout(scaled_clip, duration=1)
8text_line = TextClip(translator.translate(pr.title, dest=lan).text,
9                     font="Segoe-UI", fontsize=40, color="white")
```
06 · Distribute
Firebase Storage + FCM · 45 % less production time
Output MP4 and the ten MP3s go to Firebase Storage under gs://synth-ai-envoys.appspot.com/. Each new DescriptiveContent triggers an FCM push to every user subscribed to that ministry — title: "New Press Release", body and thumbnail derived from the record.
Net effect on the production team: ~30 videos a day, ten languages each, with the human only intervening for moderation. The 45 % time reduction and 35 % throughput lift in the project description come from this stage replacing all of the previous manual work.
07 · Try it
Run the six stages in your browser
Paste a press-release-style paragraph below. The pipeline runs the same shape as the Flask service — ingest, summarise (paragraph-split here, LLM in prod), translate to all 10 languages in parallel via /api/translate, speak the chosen language via /api/tts (gTTS upstream), compose a Canvas slideshow with the same zoom + fade + title overlay MoviePy produced, then a simulated upload.

architecture01 · The constraint01 / 07

stack

Built with.

FlaskPythonAIPrompt EngineeringImage Management

nextCollege Attendance AppDeveloped and deployed a Flutter frontend and Node.js backend application on DigitalOcean for tracking college attendance.Case study →

Want the story behind it?I answer email

Senior backend and platform roles, hard systems problems, or an argument about thread pools.

Open a connection Recruiter mode