Daiva AI Engine
An AI that actually understands documents.
Four pillars, one engine. Built in-house, hosted in your region, available via API or directly inside every screen.
Document Brain
Multimodal LLM fine-tuned on 500M+ enterprise documents. Understands layout, tables and handwriting.
Extraction Pipelines
Key-value, table and entity extraction with confidence scoring and human-in-the-loop review.
Generative Drafting
Summarize, rewrite, translate and redact — grounded in your repository, never hallucinated.
Conversational Search
Chat with your repository — citations, page snippets and source links on every answer.
Capabilities at a glance
Vision OCR/ICR
Print + handwriting in 150+ languages with edge-crop, glare removal and multi-page support.
On-the-fly translation
Translate any document in-place while preserving formatting and tables.
PII detection
Auto-detect and redact 40+ PII classes — names, IDs, payment data, medical codes.
Vector + keyword
Hybrid retrieval — best of semantic search and BM25 with custom re-rankers.
Model registry
A/B versions, per-tenant pinning and rollback in a single click. Choose accuracy vs. latency per workload.
| Model | Size | Use case | Latency |
|---|---|---|---|
| daiva-doc-base | 8B | Standard ingestion · classification · KV extraction | ≈ 250ms / page |
| daiva-doc-pro | 32B | Complex layouts · handwriting · multi-language | ≈ 600ms / page |
| daiva-chat-2 | 70B | Conversational search · drafting · summarization | ≈ 1.2s / answer |
| daiva-vision-hd | 12B | Vision OCR · forms · diagrams · charts | ≈ 400ms / page |
Your data, your region
All inference runs in your chosen region. Nothing leaves the EU / US / IN boundary.
Never trained on you
Customer data is never used to train base models. Period.
Grounded outputs
Every generative answer cites the source pages it was built from.