Up to 99.9% accuracy in production.
Self-learning. Built for receipts, not lab demos.

Most modern OCR engines claim 95–99% accuracy on header fields (merchant, total, date) — but those numbers come from clean test receipts. REME's 99.9% accuracy holds across production receipts including faded thermal paper, mobile photos, multi-language vendors, and crumpled receipts from real travel scenarios. The accuracy is also self-learning: it starts at ~96% on day one and reaches 99.9% over the first 90 days as the model learns your company's specific vendors, categories, and tax patterns.

REME maintains a per-company OCR model that captures your specific patterns. After ~50 submissions from a specific vendor, the model knows that vendor's typical amounts, categories, and tax breakdowns — even if a future receipt is partially unreadable. After 200+ corrections by your finance team, the model learns your specific category mappings and stops making the same mistakes. By month three, your model is materially better than it was on day one. This is different from generic LLM-based OCR that doesn't retain company-specific patterns.

40+ structured fields across four categories: (1) Transaction fields — total, subtotal, currency, tip, discount, payment method, date, time, receipt ID. (2) Merchant fields — name, address, tax registration, category, country, phone, format type. (3) Tax fields — GST, VAT, sales tax, service charge, tax registration validation, recoverable amount, multi-component breakdowns. (4) Line-item fields — description, quantity, unit price, item-level discount, item-level tax, category, total per line. Most legacy OCR engines extract 5–10 header fields; REME extracts the 40+ fields you'd actually want for proper accounting.

11 languages natively: English, Mandarin (Simplified + Traditional), Bahasa Indonesia, Bahasa Melayu, Hindi, Tamil, Thai, Vietnamese, Arabic, French, German. More languages added quarterly based on customer geographic distribution. Native support means the OCR engine understands receipt structure in that language — not just character recognition. A Mandarin-language receipt extracts merchant, amount, and tax breakdowns correctly; it doesn't just produce a text dump.

8 native tax engines: IRAS (Singapore), GST (India + Australia), VAT (UK + EU member states), IRS (US), HMRC (UK supplementary), FTA (UAE + GCC). Multi-component breakdowns supported (e.g., CGST + SGST + IGST for India, where GST has multiple components). The engine identifies the jurisdiction from the receipt — typically from the merchant's tax registration number or country of issue — and applies the correct tax rules automatically. For multi-entity companies operating across jurisdictions, this is one of the strongest reasons teams pick REME over geographically-concentrated tools.

1.0–2.5 seconds end-to-end on the typical receipt. The five-stage pipeline (pre-process, recognize, understand structure, extract & validate, learn) runs sequentially, but each stage is optimized for sub-second performance. The employee sees the extracted data appear in their WhatsApp conversation before they put their phone down. The submitter can correct any field with a follow-up message ('the date is March 14, not 4') and the model learns from that correction.

REME's pipeline assigns a confidence score to every extracted field (0–100%). If confidence falls below the configured threshold (default: 95%), the field is flagged for human review — either by the submitter ('we extracted SGD 7.63 — please confirm') or by the finance team during approval. Low-confidence fields don't silently flow into your accounting system as bad data. The threshold is configurable per company; some teams require 99%+ confidence on tax fields specifically because GST/VAT recovery depends on accurate amounts.

OCR is step one in a four-layer pipeline. Once OCR extracts the structured data, it flows immediately to: (1) Policy enforcement, where 5 rule categories check the extracted data against your company's policies in 200 milliseconds. (2) Fraud detection, where 7 AI agents check for duplicates, fake receipts, merchant fabrications, and anomalous patterns. (3) Approval routing, where the engine routes the claim to the right approver based on amount, category, and project. Every layer depends on accurate OCR — which is why 99.9% accuracy isn't a vanity metric but the foundation that makes the rest of the platform reliable.