AI Viral Moment Detection

def transcribe_video(video_path):
    from faster_whisper import WhisperModel
    
    # Run on CPU with INT8 quantization for speed
    model = WhisperModel("base", device="cpu", compute_type="int8")
    
    segments, info = model.transcribe(video_path, word_timestamps=True)
    
    transcript_segments = []
    for segment in segments:
        seg_dict = {
            'text': segment.text,
            'start': segment.start,
            'end': segment.end,
            'words': [
                {
                    'word': word.word,
                    'start': word.start,
                    'end': word.end,
                    'probability': word.probability
                }
                for word in segment.words
            ]
        }
        transcript_segments.append(seg_dict)
        
    return {
        'text': full_text.strip(),
        'segments': transcript_segments,
        'language': info.language
    }

GEMINI_PROMPT_TEMPLATE = """
You are a senior short-form video editor. Read the ENTIRE transcript and word-level timestamps to choose the 3–15 MOST VIRAL moments for TikTok/IG Reels/YouTube Shorts. Each clip must be between 15 and 60 seconds long.

⚠️ FFMPEG TIME CONTRACT — STRICT REQUIREMENTS:
- Return timestamps in ABSOLUTE SECONDS from the start of the video (usable in: ffmpeg -ss <start> -to <end> -i <input> ...).
- Only NUMBERS with decimal point, up to 3 decimals (examples: 0, 1.250, 17.350).
- Ensure 0 ≤ start < end ≤ VIDEO_DURATION_SECONDS.
- Each clip between 15 and 60 s (inclusive).
- Prefer starting 0.2–0.4 s BEFORE the hook and ending 0.2–0.4 s AFTER the payoff.
- Use silence moments for natural cuts; never cut in the middle of a word or phrase.

VIDEO_DURATION_SECONDS: {video_duration}

TRANSCRIPT_TEXT (raw):
{transcript_text}

WORDS_JSON (array of {{w, s, e}} where s/e are seconds):
{words_json}

STRICT EXCLUSIONS:
- No generic intros/outros or purely sponsorship segments unless they contain the hook.
- No clips < 15 s or > 60 s.

OUTPUT — RETURN ONLY VALID JSON:
{{
  "shorts": [
    {{
      "start": <number in seconds, e.g., 12.340>,
      "end": <number in seconds, e.g., 37.900>,
      "video_description_for_tiktok": "<description for TikTok oriented to get views>",
      "video_description_for_instagram": "<description for Instagram oriented to get views>",
      "video_title_for_youtube_short": "<title for YouTube Short oriented to get views 100 chars max>",
      "viral_hook_text": "<SHORT punchy text overlay (max 10 words). MUST BE IN THE SAME LANGUAGE AS THE VIDEO TRANSCRIPT>"
    }}
  ]
}}
"""

try:
    usage = response.usage_metadata
    if usage:
        # Gemini 2.5 Flash Pricing (Dec 2025)
        input_price_per_million = 0.10   # $0.10 per 1M tokens
        output_price_per_million = 0.40  # $0.40 per 1M tokens
        
        prompt_tokens = usage.prompt_token_count
        output_tokens = usage.candidates_token_count
        
        input_cost = (prompt_tokens / 1_000_000) * input_price_per_million
        output_cost = (output_tokens / 1_000_000) * output_price_per_million
        total_cost = input_cost + output_cost
        
        print(f"💰 Token Usage ({model_name}):")
        print(f"   - Input Tokens: {prompt_tokens} (${input_cost:.6f})")
        print(f"   - Output Tokens: {output_tokens} (${output_cost:.6f})")
        print(f"   - Total Estimated Cost: ${total_cost:.6f}")
except Exception as e:
    print(f"⚠️ Could not calculate cost: {e}")

curl -X POST https://your-server.com/api/process \
  -H "X-Gemini-Key: YOUR_GEMINI_API_KEY" \
  -F "url=https://youtube.com/watch?v=VIDEO_ID"

python main.py -i video.mp4 -o output.mp4 --skip-analysis