blog.backLink
ยท7 minintelligenceproductivityNLP
Automatic Insight Extraction: Decisions, Tasks, and Deadlines
After a one-hour meeting, you have thousands of words. But what actually matters? The three decisions made, five action items assigned, and the deadline mentioned in passing. dijin extracts these automatically.
5 categories
Insight Types
99 supported
Languages
6
Pipeline Stages
100% on-device
Processing
What Gets Extracted
| Insight Type | Example | What Gets Captured |
|---|---|---|
| Decisions | "We decided to go with Option B" | Participants, context, timestamp |
| Action Items | "Ayse will prepare the report by Friday" | Assignee, deadline, task description |
| Calendar Events | "Let's meet next Tuesday at 3pm" | Date/time parsed across 99 languages |
| Key Topics | Important subjects discussed | Weighted by frequency and context |
| People Mentioned | Names and references | Linked to the knowledge graph |
The 6-Stage FrameCompiler Pipeline
1
Signal Detection
Scans for temporal markers, imperative verbs, and decision language in the transcript text.
2
Entity Extraction
Identifies people, organizations, dates, and locations mentioned in the conversation.
3
Relationship Mapping
Links entities together โ who assigned what to whom, who committed to which deadline.
4
Classification
Categorizes each insight as decision, action item, calendar event, or topic.
5
Confidence Scoring
Low-confidence items go to a quarantine queue for human review before being committed.
6
Structured Output
Clean, typed data written to your local database โ ready for search and retrieval.
Calendar Detection Across 99 Languages
๐ก
The CalendarDetectionService supports 99 languages with a 7-stage parsing pipeline. "Yarin saat 10'da" (Turkish), "ๆๆฅใฎ10ๆใซ" (Japanese), and "tomorrow at 10am" (English) all resolve to the same absolute datetime.
Calendar Detection Pipeline
Raw Text: "Let's meet tomorrow at 10am"
Language Detection โโโถ English
Temporal Signal โโโถ "tomorrow" = relative
Time Extraction โโโถ "10am" = 10:00
Absolute Resolution โโโถ 2026-03-23T10:00:00
Validation โโโถ โ Valid future datetime```
Everything On-Device
๐
The entire pipeline runs locally using on-device language model. Your insights are generated and stored on your device. The cloud never sees your meeting content.