--- schema: 1 type: video title: From Raw Documents to AI-Ready Data — Leo Platzer & Jeff Koss description: >- A chatbot pilot on 40 hand-picked files works great. Point it at 80,000 SharePoint documents and it falls apart. Jeff Koss and Leo Platzer, co-founder and former CTO of Deasy Labs (acquired by Collibra), walk through taking an unstructured-data AI project from stalled POC to production. Using a manufacturer building a legal-ops chatbot over 80,000 SharePoint files, they demo AI-assisted taxonomies and metadata tagging with evidence and confidence scores, sensitive-data detection, and quality checks for duplicates, conflicts and freshness. Data slices refresh on a schedule. Leo then shows why this matters: when 30% of a corpus is stale or duplicated, up to 80% of an agent's retrieved context can be useless, and cleaning it roughly doubled recall in multi-hop RAG evals. He finishes by generating a context file for coding agents with the SDK. In this talk: • Why document data quality means duplicates, conflicts and freshness • AI-assisted taxonomies and metadata tagging with evidence and confidence • Finding and filtering sensitive data before it reaches a chatbot • How stale, duplicated data crowds out good context, and what cleaning it does to recall SPEAKERS Leo Platzer, Co-founder & former CTO, Deasy Labs (Collibra) LinkedIn: https://www.linkedin.com/in/leonardplatzer/ Jeff Koss, Staff Customer Engineer, Deasy Labs (Collibra) LINKS Collibra unstructured data for AI: https://www.collibra.com/use-cases/solution/unstructured-data-for-ai Collibra acquires Deasy Labs: https://www.collibra.com/company/newsroom/press-releases/collibra-acquires-deasy-labs Deasy Labs: https://www.deasylabs.com Collibra: https://www.collibra.com/ CHAPTERS 0:00 Intro 0:18 Deasy Labs and Collibra 1:01 From stalled POC to production 1:11 Scenario: a legal-ops chatbot 2:11 Who needs what 3:25 40 files works 3:55 80,000 files doesn't 4:25 Data quality for documents 5:05 The symptoms 6:35 Demo: tagging with a taxonomy 9:04 Metadata with evidence 10:04 Finding sensitive data 10:54 Duplicates, conflicts, freshness 11:54 Data slices and suggested taxonomies 13:33 Keeping data fresh 14:13 Results: four months to days 15:00 Why duplicates poison retrieval 16:54 Stale data can fill 80% of context 17:39 Results: about 2x recall 18:29 Demo: a context file for coding agents 20:36 Wrap-up Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #RAG #DataQuality #AIEngineer canonical: 'https://liquidgold.cc/watch/wzWNYDY7toc' markdown: 'https://liquidgold.cc/watch/wzWNYDY7toc.md' llms: 'https://liquidgold.cc/watch/wzWNYDY7toc/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:19:50.677Z' language: null children: [] videoId: wzWNYDY7toc durationSeconds: 1297 publishedAt: '2026-10-05T22:30:12.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 210 endSeconds: 221 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=210' heldReason: rights - ordinal: 1 text: null startSeconds: 221 endSeconds: 244 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=221' heldReason: rights - ordinal: 2 text: null startSeconds: 278 endSeconds: 312 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=278' heldReason: rights - ordinal: 3 text: null startSeconds: 416 endSeconds: 544 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=416' heldReason: rights - ordinal: 4 text: null startSeconds: 556 endSeconds: 604 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=556' heldReason: rights - ordinal: 5 text: null startSeconds: 604 endSeconds: 615 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=604' heldReason: rights - ordinal: 6 text: null startSeconds: 663 endSeconds: 676 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=663' heldReason: rights - ordinal: 7 text: null startSeconds: 723 endSeconds: 735 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=723' heldReason: rights - ordinal: 8 text: null startSeconds: 816 endSeconds: 827 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=816' heldReason: rights - ordinal: 9 text: null startSeconds: 903 endSeconds: 914 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=903' heldReason: rights - ordinal: 10 text: null startSeconds: 1017 endSeconds: 1029 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=1017' heldReason: rights - ordinal: 11 text: null startSeconds: 1062 endSeconds: 1075 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=1062' heldReason: rights - ordinal: 12 text: null startSeconds: 1124 endSeconds: 1242 watch: 'https://liquidgold.cc/watch/wzWNYDY7toc?t=1124' heldReason: rights --- # From Raw Documents to AI-Ready Data — Leo Platzer & Jeff Koss A chatbot pilot on 40 hand-picked files works great. Point it at 80,000 SharePoint documents and it falls apart. Jeff Koss and Leo Platzer, co-founder and former CTO of Deasy Labs (acquired by Collibra), walk through taking an unstructured-data AI project from stalled POC to production. Using a manufacturer building a legal-ops chatbot over 80,000 SharePoint files, they demo AI-assisted taxonomies and metadata tagging with evidence and confidence scores, sensitive-data detection, and quality checks for duplicates, conflicts and freshness. Data slices refresh on a schedule. Leo then shows why this matters: when 30% of a corpus is stale or duplicated, up to 80% of an agent's retrieved context can be useless, and cleaning it roughly doubled recall in multi-hop RAG evals. He finishes by generating a context file for coding agents with the SDK. In this talk: • Why document data quality means duplicates, conflicts and freshness • AI-assisted taxonomies and metadata tagging with evidence and confidence • Finding and filtering sensitive data before it reaches a chatbot • How stale, duplicated data crowds out good context, and what cleaning it does to recall SPEAKERS Leo Platzer, Co-founder & former CTO, Deasy Labs (Collibra) LinkedIn: https://www.linkedin.com/in/leonardplatzer/ Jeff Koss, Staff Customer Engineer, Deasy Labs (Collibra) LINKS Collibra unstructured data for AI: https://www.collibra.com/use-cases/solution/unstructured-data-for-ai Collibra acquires Deasy Labs: https://www.collibra.com/company/newsroom/press-releases/collibra-acquires-deasy-labs Deasy Labs: https://www.deasylabs.com Collibra: https://www.collibra.com/ CHAPTERS 0:00 Intro 0:18 Deasy Labs and Collibra 1:01 From stalled POC to production 1:11 Scenario: a legal-ops chatbot 2:11 Who needs what 3:25 40 files works 3:55 80,000 files doesn't 4:25 Data quality for documents 5:05 The symptoms 6:35 Demo: tagging with a taxonomy 9:04 Metadata with evidence 10:04 Finding sensitive data 10:54 Duplicates, conflicts, freshness 11:54 Data slices and suggested taxonomies 13:33 Keeping data fresh 14:13 Results: four months to days 15:00 Why duplicates poison retrieval 16:54 Stale data can fill 80% of context 17:39 Results: about 2x recall 18:29 Demo: a context file for coding agents 20:36 Wrap-up Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #RAG #DataQuality #AIEngineer Canonical: https://liquidgold.cc/watch/wzWNYDY7toc · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:00 to 1:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:12 to 2:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:19 to 3:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (3:30 to 4:27) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (4:27 to 5:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (5:12 to 6:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (6:45 to 10:04) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (10:04 to 11:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (11:03 to 12:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (12:03 to 13:36) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (13:36 to 14:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (14:14 to 15:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (15:03 to 17:42) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (17:42 to 18:32) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (18:32 to 20:42) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (20:42 to 21:37) > Held: transcript rights do not permit this chunk's text.