•Projects that speak for themselves

Our Case Studies

Check out our portfolio to view the featured projects!

First screen

•Case study

RAG-Powered Natural Language to SQL for Accounts Payable Automation

A proof of concept using RAG architecture to enable non-technical users to interact with accounting data through natural language queries translated into secure SQL.

Industry:

Accounts Payable Automation

Location:

Australia

Duration:

3 months

Budget:

$15K

•Case study

Our Approach

We approached this engagement as a RAG-driven AI data access project, rather than a generic NL-to-SQL experiment.

Our focus was on: Designing a secure AI interface between natural language input and a financial SQL database.

Implementing Retrieval-Augmented Generation (RAG) to ground AI output in database schema, metadata, and query rules.

Prioritizing accuracy, safety, and auditability, critical for finance and accounting workflows.

Delivering a fast PoC while maintaining a clear path to production readiness.

Instead of relying on raw LLM output, we structured the system so the model reasons over retrieved schema context before generating SQL, significantly reducing errors and risk.

The Problem

For financial SaaS platforms, enabling easy data access introduces several challenges: Business users need insights but lack SQL expertise.

Direct AI-to-database interaction can lead to unsafe or incorrect queries.

Accounting data demands strict control, validation, and traceability.

Any AI solution must integrate cleanly with existing cloud infrastructure.

The core challenge was proving that AI could generate accurate, secure SQL queries from natural language without compromising data integrity.

RAG-Based Natural Language to SQL

Screen

Secure Cloud Integration

Screen

API & Interface Layer

RAG-BASED NL-TO-SQL POC

We delivered a working PoC and after MVP with the following capabilities:

KEY FEATURES IMPLEMENTED:

01
RAG-Based Natural Language to SQL
A retrieval layer supplies the LLM with database schema and table relationships, column definitions and constraints, and predefined query boundaries. This context-driven approach enabled consistent, reliable SQL generation.
02
Secure Cloud Integration
The solution integrated AWS RDS (SQL) as the data source and Azure OpenAI (GPT-4o) for language understanding and generation. Strict access controls ensured safe execution of generated queries.
03
API & Interface Layer
We built a simple interface that accepts natural language questions, generates validated SQL statements, and returns structured results for downstream use.
04
Production-Aware Architecture
Although delivered as a PoC, the system was designed with clear separation between AI, data, and execution layers, scalability and governance in mind, and readiness for future role-based access and logging.

Here's how it looks

•Projects that speak for themselves

What Our Clients Say

Trust is built on results. Here’s what our partners think about working with us.

Xedrum delivered a production-ready AI-powered application within the agreed timeline and budget, including a working release on Apple TestFlight. The team was highly proactive in resolving issues, adapting to evolving requirements, and ensuring the AI functionality aligned with real user needs. Throughout the engagement, Xedrum consistently treated the project as a priority and demonstrated a strong understanding of both technical and business concerns.

Celine LeibfriedCEO at Unscripted

Xedrum successfully implemented a robust system that allows users to upload data, execute automated jobs, and retrieve results efficiently. The team delivered all components on schedule, maintained clear communication through weekly and ad-hoc meetings, and demonstrated professionalism throughout. The final solution met both technical and operational expectations.

Tomer PeretzCTO at Osirix

Xedrum's engineers worked closely with the client's project manager, integrating seamlessly into the existing workflow. The team consistently delivered planned milestones, maintained strong communication through regular online meetings, and brought a personal, collaborative approach that made them feel like an in-house AI team rather than an external vendor.

FredericCTO at Touch2Seen

Xedrum operated as an extension of the client's internal team, taking ownership of both delivery and day-to-day execution. Their engineers quickly understood the product context, collaborated across time zones, and proactively contributed ideas to improve the solution. The team maintained consistent velocity, adapted to changing priorities, and helped move the product forward without adding management overhead.