# Simak (سِماك)

> Arabic-first agentic document understanding platform for the Saudi market.
> Users describe in plain language what they need extracted; Simak's wizard
> interviews them, designs a JSON schema, and deploys an extractor behind an
> API key. Built by AI researchers using purpose-built methods for Arabic,
> not English-first OCR with Arabic bolted on.

## Identity

- Name: Simak (Arabic: سِماك, transliterated Simāk)
- Etymology: The root س-م-ك means "to hold up, to support, to raise". Simākān are the two stars (Arcturus + Spica) named for appearing to *hold up* the sky. The product mirrors this: it gives structure to unstructured documents.
- Domain: https://simak.ai
- Headquarters: Dammam, Saudi Arabia
- Hosting region: Google Cloud me-central2 (Dammam, Saudi Arabia)
- Compliance: PDPL (Saudi Personal Data Protection Law). No training on customer data, ever.

## What Simak does

- End-to-end document workflow: input (WhatsApp, email, file upload, API push, mobile camera) → preprocess (deskew, denoise, split) → classify → extract → output (JSON, Google Sheets, Excel, webhooks, ERP/accounting connectors).
- Schema designed through conversation (no code required), with a developer surface (prompt registry, schema overrides, model client) underneath. Non-technical users and developers ship from the same platform.
- Arabic-first by design: tuned decoding profile, source grounding, two-pass verification on critical fields, self-consistency sampling, constraint-as-contract validation, bidi-aware rendering.

## Differentiators

1. Fully agentic and no-code by default, with a developer surface underneath.
2. Built by AI researchers using purpose-built Arabic methods.
3. Accuracy-first, edge-case ready. North-star: ≤1% error on critical fields in strict mode.
4. KSA data residency by construction (me-central2, Dammam). PDPL-compliant retention defaults.
5. Ready-made templates for the Saudi document stack: ZATCA e-invoices, financial statements (balance sheets, income statements, cash flow), GOSI certificates, commercial registration (CR), VAT registration and returns.
6. End-to-end document workflow, not just an extraction box. Composable input, preprocessing, classification, extraction, and output stages.

## Quality stance

- No fuzzy 0-to-1 confidence numbers anywhere.
- Every field carries a deterministic `validity` block: constraint check, source grounding, two-pass agreement, self-consistency.
- A `review_strictness` setting (lenient / normal / strict) controls which failures flag a job for human review.
- A `quality_mode` setting (fast / balanced / strict) is the cost/accuracy lever.
- In-product review queue for flagged jobs. Human corrections fix the single job and do NOT become training data.

## Audience

Saudi businesses in fintech, legal, healthcare, government, and insurance who handle Arabic and bilingual Arabic-English documents and care about defensible accuracy on KYC, KYB, contracts, medical records, and government filings.

## Sections

- [Home](https://simak.ai/): Hero introduction
- [Platform](https://simak.ai/#platform): Five-stage agentic pipeline (input, preprocess, classify, extract, output)
- [Methods](https://simak.ai/#methods): Arabic-first methodology
- [Templates](https://simak.ai/#templates): KSA-tailored presets
- [Waitlist](https://simak.ai/#waitlist): Early access

## Optional

- [llms-full.txt](https://simak.ai/llms-full.txt): Full content bundle of all sections for direct ingestion.