Haier Document-Intelligence Platform

Highlights

1. Vendor-agnostic LLM layer swapped across local Ollama qwen3-vl, AWS Bedrock Qwen3-VL-235B, and Gemini, chosen per use case on measured cost, latency and quality.
2. Map-reduce pipeline for 15–20+ page legal PDFs with a context memo carried across chunked LLM calls and a global-best reduce phase.
3. Evidence-based invoice arithmetic validator with an OCR digit-confusion model: it cross-checks totals and tax and only auto-corrects when multiple independent signals agree.
4. Hybrid OCR↔LLM repair: PaddleOCR text repairs GSTIN and dirty dates the LLM misread, instead of trusting either source blindly.
5. Classify→crop→normalise preprocessing router (ResNet-18) with multithreaded page rendering.