TARUNAI Systems Engineer← All work
02 ENTERPRISE AI · CASE STUDY

Governed natural-language access to enterprise ERP information.

AJSMGPT translates supported business questions into controlled, validated Oracle ERP queries—without giving a language model unrestricted SQL access.

Role
Independent system design and development
Environment
Internal enterprise ERP workflow
Stack
FastAPI · Oracle · Qdrant · LLMs
1,574Logged question records from development and evaluation
74Approved regression cases
34Structured feedback-repair cases
<3 secTarget response time for supported approved queries

Response time is a performance target. Organisation-wide time and cost savings have not yet been formally measured.

ERP information was available, but not naturally accessible.

Employees often depend on manual ERP navigation, predefined reports, or technical SQL support to retrieve operational information.

A general-purpose model generating unrestricted SQL would create unacceptable accuracy and security risks. The real challenge was building useful natural-language access while keeping behaviour controlled, testable, and reversible.

Natural-language interaction around deterministic controls.

The model assists interpretation; verified routes and database safeguards govern execution.

  1. 01

    Business question

    A user asks for operational information in everyday language.

  2. 02

    Controlled routing

    Supported intent and parameters are matched to verified logic.

  3. 03

    Validated query

    Read-only Oracle access runs through constrained query patterns.

  4. 04

    Safe answer

    Results are formatted for the user with traceable system behaviour.

A safe answer carries evidence of the route behind it.

This demonstration contains no client queries, supplier identities, database values, or production information.

New behaviour earns its way into production.

Publishing is a separate human-controlled action, with versions, audit history, disable support, and rollback.

  1. 01Open
  2. 02Candidate generated
  3. 03Tested
  4. 04Human approved
  5. 05Regression passed
  6. 06Ready to publish
  7. 07Published

The assistant is useful because its freedom is intentionally limited.

01

Read-only database access

The system retrieves information; it cannot modify operational ERP records.

02

Verified query routes

Supported questions use reviewed tables, joins, filters, and calculations.

03

Regression protection

Approved behaviours are retested before new versions can be published.

04

Human release control

Testing does not equal publishing; an authorized person makes the final release decision.

Independent engineering across AI, enterprise data, evaluation, and release.

I designed and developed the system independently. Domain users supported business-rule validation and query-correctness feedback.

01FastAPI service and Oracle integration
02Deterministic query routers and metadata retrieval
03Read-only safety and validation controls
04Evaluation and regression pipeline
05Feedback-learning and review interface
06Versioned publishing, audit, and rollback workflow

Controlled scope is part of the design.

Need trustworthy AI access to complex business information?

I’m available for selected enterprise-assistant, retrieval, and governed-AI projects.

Discuss a project ↗Back to portfolio