Infonexs

Current PoC and Foundations

Current PoC: ERP Text-to-SQL (In Development)

Technical Overview: A plain-language query layer over a sample ERP database covering receivables, inventory, sales and purchasing.

Architecture: Read-only database role, business glossary mapped to tables and columns, model-generated SQL parsed and checked against the schema before execution, full query and result logging.

Capability to Prove: Execution accuracy and latency on a 200-question benchmark, first with cloud models, then with open-source models on local hardware.

Status: Benchmark results will be published on this page.

“Which receivables over $10k are overdue?”GENERATED SQLSELECT customer, amount_due FROM ar_invoicesWHERE due_date < CURRENT_DATE AND amount_due > 10000 LIMIT 100;Schema check · read-only role · row limitCustomerAmount dueCustomer A$24,300Customer B$18,950
Policy documentspgvectorretrievalDRAFT[1][2]Advisor approves → PublishEvery draft links back to the passages it was built from.

Foundation 1: Complex Document RAG & Source Attribution

Technical Overview: A content engine deployed for an education advisory service, indexing frequently changing policy documents and institutional data.

Architecture: PostgreSQL + pgvector retrieval, drafts linked to their source passages, and advisor approval before publishing.

Capability Proven: Accurate retrieval over unstructured documents, with every output traceable to its source.

Carries into AIHub as: The policy and document layer for procurement rules, SOPs and accounting policies.

Foundation 2: Rule-Validated Generation

Technical Overview: An adaptive question generator deployed for a language-training platform.

Architecture: Model output checked against an explicit rule base and knowledge graph before it reaches a learner.

Capability Proven: Constraining model output with rules defined outside the model.

Carries into AIHub as: The SQL validation layer, where generated queries must pass schema and policy checks before they run.

Model outputquestion draftRule base +knowledge graph checkFail → regeneratePass → deliveredto the learnerRules live outside the model, so they can be audited and changed.
On-deviceprocessingruns firstCloud service(Cloudflare)heavier jobsresultsLight work stays on the device; the cloud absorbs peaks.

Foundation 3: Hybrid On-Device and Cloud Processing

Technical Overview: An image-processing pipeline for a consumer mobile app in production.

Architecture: Processing runs on the device first; heavier jobs route to a Cloudflare-based cloud service.

Capability Proven: Splitting AI workloads between the edge and the cloud to control latency and cost.

Carries into AIHub as: Routing between local models on customer hardware and cloud models.

Our Design Principle

The database stays the source of truth. AI writes the query and drafts the action; your systems compute the result and your people approve the change.— Infonexs