We built a real-time pipeline that transcribes meeting audio and uses a natural language processing layer to extract tasks, deadlines, and assignees.
To filter out casual chatter and off-topic noise, we use semantic similarity algorithms to compare the overall meeting context against potential action items. Only statements that highly align with actual tasks are classified and automatically dispatched to the team's project board.How we built it
ChalleLatency Bottleneck: Sending massive live transcripts to the AI model caused a 10-second response delay. We resolved this by implementing a streaming sliding window approach, dropping the latency to under 1.5 seconds.
Speaker Diarization: To prevent the system from assigning tasks to the wrong people, we integrated voice-profile mapping to ensure that when a speaker says "I will handle the deployment," the ticket is assigned to their specific account.nges we ran into
High-Accuracy Extraction: We achieved an accuracy rate of over 90% in extracting action items without cluttering the target project boards with meeting noise.
Seamless API Integration: We successfully built automated sync pipelines that turn spoken words into fully populated Jira tickets, Trello cards, and Notion database items in near real-timlishments that we're proud of
Structured Model Orchestration: We mastered prompt engineering techniques and structured outputs to ensure our backend always sends clean, predictable data to external APIs.
Real-Time Data Streaming: We gained deep hands-on experience managing low-latency connections and streaming audio buffers under high-volume conditions.
Cross-Language Support: We plan to expand our translation and transcription capabilities to support multilingual meetings seamlessly.
Deep Calendar Integration: The next step is integrating with Google Calendar and Outlook to automatically draft follow-up invitations based on the deadlines agreed upon during the meeting.
Log in or sign up for Devpost to join the conversation.