Ask your data in Arabic
Ask your enterprise data in Arabic or English. Bilytica NL2SQL turns plain questions into governed, secure answers, in seconds, inside Saudi Arabia.
Most Saudi teams wait days for a dashboard answer. We close that gap. Ask your warehouse, ERP, or CRM a question in plain Arabic or English, and get a governed answer pulled through a certified semantic layer, with SSO identity, PII masking, and a full audit trail built in.
What We Do
Ask your data in Arabic capabilities
- 01
Arabic + English NL2SQL
Bilingual query parsing tuned for Saudi-MSA business vocabulary، أرباح، تسويات، مطالبات، مخزون، not machine-translated English prompts.
- 02
Certified Semantic Layer of Data
We build and certify metrics, dimensions, hierarchies, and joins on dbt Semantic Layer, Cube, LookML, MicroStrategy, or AtScale, so 'revenue', 'active customer', or 'NPL ratio' resolves to one governed definition, every time.
- 03
Complete Enterprise Security Layer
Seven-layer defence: Saudi SSO identity propagation, row- + column-level RBAC, dynamic PII/PHI masking, prompt-injection & jailbreak filters, safe-SQL execution sandbox (read-only, no DDL/DML), rate-limits per user & metric, and immutable audit of every prompt, generated SQL, and result row.
- 04
ERP / CRM / Warehouse Connectors
Pre-built adapters for SAP S/4HANA, Oracle Fusion, Odoo, Salesforce, Snowflake, Databricks, Azure Synapse, BigQuery, and PostgreSQL / SQL Server, reading through the semantic layer, never around it.
- 05
Hallucination & Accuracy Harness
Execution-accuracy and exact-match evaluation against a curated Saudi question set, with weekly regression runs before any model or schema change ships.
- 06
In-country / HUMAIN Hosting
Deployable on HUMAIN, STC Cloud, Oracle Jeddah, or on-prem GPU, PDPL data-residency by design, with no prompts or query results leaving Saudi Arabia.
Our Approach
From your data to a working result
- 01
Question Inventory
Shadow finance, ops, and executive teams for two weeks to capture the 200 questions they actually ask, in Arabic and English.
- 02
Semantic Layer Hardening
Certify metrics, dimensions, and joins in dbt / Cube / LookML so every generated query resolves against one governed definition of truth.
- 03
Bot Architecture
Design retrieval (schema + few-shot), prompt guardrails, safe-SQL execution sandbox, and cite-your-tables output, model-agnostic (Bilytica, GPT-5, Gemini, ALLaM).
- 04
Arabic Tuning
Fine-tune on Saudi business phrasing, MSA number/date formats, hijri/gregorian resolution, and dialect-to-MSA normalization.
- 05
Evaluation Harness
Build a golden question set with expected SQL and results; run execution-accuracy + exact-match tests before every release.
- 06
PDPL & RBAC
Wire the bot to Saudi SSO, propagate identity into row- and column-level policies, mask PII, and log every query for full auditability.
- 07
Operate & Improve
Weekly review of unanswered questions, failed SQL, and low-confidence answers, feeding the semantic layer and few-shot library.
Industry Applications
Where it makes a difference
Technologies & Models We Use
Enterprise-grade tools and frameworks powering our solutions.
Snowflake
Data Cloud
Databricks
Lakehouse
Microsoft Fabric
Analytics Suite
Power BI
Visualization
Tableau
Visualization
Azure
Cloud
AWS
Cloud
Python
Engineering
Questions buyers ask us
Straight answers before you commit
What is an NL2SQL bot and how is it different from a chatbot?
An NL2SQL bot converts a natural-language question, in Arabic or English, into a governed SQL query executed against your ERP, CRM, and warehouse, then returns a chart or table. Unlike a generic chatbot, every query passes through a certified semantic layer and a seven-layer enterprise security stack (SSO, row/column RBAC, PII masking, prompt-injection guards, safe-SQL sandbox, rate limits, immutable audit).
Does Bilytica NL2SQL support Arabic questions natively?
Yes. Bilytica NL2SQL bots are Arabic-first, tuned on Saudi Modern Standard Arabic business vocabulary، أرباح، تسويات، مطالبات، مخزون، with Hijri/Gregorian date resolution and dialect-to-MSA normalization. English is fully supported in parallel.
How do you prevent hallucinated or unsafe SQL?
Every generated query resolves against a certified semantic layer (dbt / Cube / LookML), runs inside a read-only safe-SQL sandbox with no DDL/DML, and is validated by an execution-accuracy and exact-match evaluation harness against a curated Saudi golden question set before every release.
Is the NL2SQL bot compliant with PDPL and SAMA?
Yes. Deployments are in-country on HUMAIN, STC Cloud, Oracle Jeddah, or on-prem GPU. Saudi SSO identity propagates into row- and column-level policies, PII/PHI is masked at the SQL boundary before the LLM sees a row, and every prompt, generated SQL, and returned row is written to an immutable audit log for SAMA and PDPL review.
Which data sources can the bot query?
Pre-built connectors cover SAP S/4HANA, Oracle Fusion, Microsoft Dynamics, Odoo, Salesforce, Snowflake, Databricks, Azure Synapse, BigQuery, Redshift, PostgreSQL, SQL Server, Oracle DB, and MySQL, always reading through the semantic layer, never around it.
How long does a Bilytica NL2SQL deployment take?
A production pilot is scoped with you and phased: question inventory, semantic-layer hardening, bot architecture with Arabic tuning, evaluation harness, and PDPL/RBAC wiring, followed by weekly operate-and-improve cycles.
Ready to Explore Ask your data in Arabic?
15 minutes. One strategy call. A customized roadmap tailored to your business.
Try the NL2SQL Bot: Safe Query, Semantic Layer, Full Audit
Pick a sample question or type your own. See the SQL generated through the semantic layer, a mocked result, and the security layers applied before any row is returned.
How the bot read the question
…
Mocked result
Generated SQL (read-only, via semantic layer)
-- …Security layers applied to this query
- Read-only sandbox, no DDL/DML permitted
- Row-level policy filters by caller's business unit
- PII columns (national_id, iban) auto-masked before return
This is an interactive preview, not connected to a live warehouse. Book a live demo on your data.
NL2SQL Bots Security Stack: How the Defences Are Layered
From an Arabic or English user question to an audited row in the warehouse, under full governance.
- 01
Saudi SSO & Identity Propagation
Nafath / Absher / Azure AD identity flows into every generated SQL query, no shared service accounts.
- 02
Row & Column-Level RBAC
Fine-grained policies restrict rows by region, branch, or business unit and mask columns by role.
- 03
Certified Semantic Layer
dbt / Cube / LookML governs every metric so 'revenue' or 'NPL ratio' resolves to one definition.
- 04
Dynamic PII / PHI Masking
National IDs, IBANs, and patient data masked at the SQL boundary before the LLM ever sees a row.
- 05
Prompt-Injection & Jailbreak Guards
Multi-stage filters, allow-listed operations, and role-scoped intents block malicious prompt patterns.
- 06
Safe-SQL Execution Sandbox
Read-only, no DDL/DML, cost caps, and per-user rate limits, the bot cannot mutate your warehouse.
- 07
Immutable Audit & PDPL Log
Every prompt, generated SQL, and returned row is logged in-country for SDAIA / SAMA / PDPL audit.