Document intelligence · Back office · Finance · Gov
Paperwork in Clean data out
We build end-to-end document intelligence pipelines for scanned invoices, forms, and PDFs in Vietnamese and Japanese—combining layout analysis, OCR, VLM extraction, deterministic schema rules, and a targeted human review queue.
- Problem
- Scanned invoices and forms in Vietnamese and Japanese, still typed in by hand.
- Proven
- VI/JA OCR in production: accuracy 94% → 98%, latency ~40% lower.
HÓA ĐƠN GIÁ TRỊ GIA TĂNG
請求書 · VAT invoice
Đơn vị bán: Công ty TNHH Minh An
Mã số thuế: 0312 457 889
Ngày 18 tháng 11 năm 2026
| Hàng hóa | SL | Thành tiền |
|---|---|---|
| Thùng carton 5 lớp | 1,200 | 21,600,000 |
| Băng keo OPP 100y | 400 | 6,400,000 |
| Pallet nhựa 1200 | 60 | 13,200,000 |
| Phí vận chuyển | 1 | 3,818,519 |
Thuế GTGT 8%: 3,601,481
Tổng cộng: 48,620,000 ₫
(Ký, ghi rõ họ tên) — 署名
Extracted
schema: vat_invoice.v3
Architecture
Built for production Not a slide deck
Under the hood
data → product
VI/JA OCR in production: accuracy 94% → 98%, latency ~40% lower.
What you get
Concrete outputs Measured on your data
Schema-valid JSON
PaddleOCR layout + OCR and VLM extraction validated against your target schema with deterministic rules (totals, tax, dates) and an LLM judge.
Human review only where unsure
Pages that pass schema checks flow straight through; uncertain fields route to a Label Studio queue and Next.js review app instead of slowing down the whole batch.
RAG-ready index
Extracted documents are chunked and indexed into Qdrant for cited search and Q&A, backed by Triton + TensorRT GPU inference, Redis, and PostgreSQL.
How we deliver
Small team Short loops
- 01
Start from your real data
A week with your documents, cameras or logs. We agree on one number that means “working” and freeze a test set before we train anything.
- 02
Prototype behind a gate
Baselines first, then models and agents. Every change runs against the frozen set, so progress is a chart, not an opinion.
- 03
Harden for production
Quantize, serve, trace and cost it. Retraining, rollback and monitoring are built in, not bolted on.
- 04
Hand over, stay close
Your team gets the code, the decision records and a walkthrough. We stay on call while it settles.
FAQ
Common questions Direct answers
- What languages and document formats does the Document AI pipeline handle?
- The pipeline processes scanned invoices, forms, and PDFs ingested from S3, MinIO, or a FastAPI upload API, with production OCR tuned for Vietnamese and Japanese layouts as well as English documents.
- How do you prevent noisy scans or OCR errors from reaching downstream systems?
- Every extracted record runs through deterministic validation rules—checking fields like totals, tax, and dates—alongside an LLM judge. Only records that fail validation or fall below confidence thresholds are sent to a Label Studio human review queue.
- How is Document AI deployed and served in production?
- Layout, OCR, and extraction models run on Triton Inference Server with TensorRT on GPU infrastructure, orchestrated with Redis and PostgreSQL, with optional Qdrant indexing and a Next.js review interface.
- How do we start a Document AI engagement with Untern?
- We start with a week on your real documents, agree on one metric that defines working, and freeze a test set before training or tuning anything. You can begin with a Proof Sprint for a runnable prototype and clear go/no-go numbers before committing to full production.
Related solutions
Have a problem AI should solve?
Tell us what the data looks like and what “working” would mean for you. One of the founders will reply.