Skip to main content
KommtKevinOnline uses advanced AI technology to analyze stream endings and predict when Papaplatte will go live next. The system processes the final minutes of each stream to extract Kevin’s statements about his future streaming plans.

The prediction pipeline

The AI-powered prediction system works through a multi-stage process:
1

VOD audio extraction

When a stream ends, the system retrieves the VOD (Video on Demand) from Twitch and extracts the audio from the final minutes of the stream.
2

Whisper speech-to-text

The audio is processed using OpenAI’s Whisper model, which transcribes Kevin’s spoken words into text with high accuracy. This creates a complete transcript of the stream ending.
3

ChatGPT intent classification

The transcript is analyzed by ChatGPT, which classifies whether Kevin expressed intent to stream again. The AI looks for phrases like “see you tomorrow at 3pm” or “going live again on Monday.”
4

Date extraction and storage

If streaming intent is detected, the AI extracts specific dates and times mentioned by Kevin. These predictions are stored in the database as onlineIntendDate values.

Speech-to-text with Whisper

Whisper is OpenAI’s state-of-the-art speech recognition model that converts Kevin’s audio into accurate text transcripts. The system specifically analyzes the last few minutes of each stream, where streamers typically announce their next streaming schedule. Key benefits:
  • High accuracy even with background noise or music
  • Handles German language and colloquialisms naturally
  • Creates timestamped segments for synchronized playback

Intent classification with ChatGPT

Once the transcript is generated, ChatGPT analyzes the text to determine:
  • Does Kevin plan to stream again? (Yes/No classification)
  • When does he plan to stream? (Time and date extraction)
  • How certain is the commitment? (Confidence assessment)
The AI understands natural language patterns like:
  • “Wir sehen uns morgen um 15 Uhr” (See you tomorrow at 3pm)
  • “Nächster Stream am Montag” (Next stream on Monday)
  • Relative time references (“tomorrow,” “in two days,” etc.)
Prediction accuracy depends on Kevin clearly stating his plans at the end of the stream. The AI cannot predict streams that weren’t announced or spontaneous streaming sessions.

Prediction logic

The app uses smart logic to determine if Kevin will stream today:
  • If there was a stream today and Kevin said he’d stream tomorrow, the prediction applies to the next day
  • If the last stream was yesterday or earlier and Kevin mentioned streaming today, that prediction is shown
  • Multiple predicted times can be stored and displayed in chronological order

Viewing predictions

When the AI detects streaming intent, the homepage displays:
  • “Ja” (Yes) in large text if a stream is predicted for today
  • Predicted time - “Kevin plant um 15:00 Uhr live auf Twitch zu sein” (Kevin plans to go live on Twitch at 3:00pm)
  • Late stream warning - If the predicted time has passed but Kevin isn’t live yet, a humorous message appears
All predictions are based on Kevin’s own statements. The AI doesn’t make assumptions—it only reports what Kevin said during his streams.

Build docs developers (and LLMs) love