CASE STUDY · Private AI & decision support

Put AI to work where sensitive information needs to stay inside your environment.

An air-gapped Hebrew-language decision-support system combining deterministic rules, a local language model and human approval.

Discuss a private AI deployment

Technical ownership: Itzhak Shaul · Architecture & hands-on development
Engagement: Selected freelance project · Client identity confidential
Core technologies: Python / FastAPI · vLLM / Ollama · Hebrew NLP · PostgreSQL / pgvector · React

Cloud-only AI is not the right fit for every environment. This system demonstrates how local models, explicit rules and human review can provide useful decision support while the workload stays inside an air-gapped deployment.

A strong fit for: Organizations with sensitive Hebrew-language material or deployment requirements that rule out external cloud-model calls.

The business challenge

The system needed to analyze Hebrew text and produce structured, reviewable decisions without calling external cloud services. A useful answer also needed traceability to the source and a way for people to approve or override it.

My responsibility

I architected and developed the local AI system, combining Hebrew NLP, a hybrid decision engine, structured outputs, a review interface and containerized deployment.

How the system fits together

Simplified workflow illustrating the delivered capabilities.

Local documents
Rules & local model
Cited decision
Human approval

What I built

Local language processing

Hebrew morphological analysis, named-entity recognition and text classification.

Rules plus a local model

A deterministic rule engine combined with a local LLM through vLLM or Ollama.

Reviewable decisions

Structured JSON with a decision, confidence score and citations from the source text.

Human oversight

An audit log, approval and override interface, and Docker Compose deployment in an air-gapped environment.

ARCHITECTURE & IMPLEMENTATION

The technology, and the work behind it.

How the documented stack supports the business workflow, from the application layer to integrations and operational controls.

Application service

Python · FastAPI

A local service coordinates text analysis and the decision workflow. The system runs without calls to external cloud services, making the deployment boundary a core architecture requirement.

Hebrew text analysis

Hebrew NLP · Morphological analysis · Entity recognition

Language processing covers morphology, named entities and classification. This prepares Hebrew text for a decision process that needs more structure than a generic chat interface.

Hybrid decision engine

Deterministic rules · Llama / Qwen / Mistral · vLLM / Ollama

A rule engine works alongside a local language model. The documented local-model stack includes Llama, Qwen and Mistral through vLLM or Ollama, combining explicit decision logic with language interpretation.

Structured and traceable output

JSON · Confidence scores · Source citations · PostgreSQL / pgvector

The decision output includes a confidence score and citations from the source text. PostgreSQL / pgvector forms part of the documented data stack, while structured results give the application a defined interface for review.

Operator review and deployment

React · Audit log · Human approval / override · Docker Compose

A review UI lets people approve or override the result. The audit trail and containerized deployment support an inspectable local workflow rather than an unreviewed model response.

Follow the workflow

An illustrative walkthrough of the capabilities described in this project.

Hebrew text is analyzed locally, with language-processing steps identifying relevant structure and entities.
The rule engine and local model produce a structured decision with confidence information and supporting source citations.
A person reviews the result and can approve or override it; the system records the decision path in its audit log.

The engineering behind the result

The model supports a decision process with explicit rules and human oversight. Source citations and structured outputs make the result inspectable rather than an opaque paragraph of generated text.

What this enables

The delivered system enables local AI-assisted analysis with a documented review path and no calls to external cloud services.

BUILD A SYSTEM AROUND YOUR BUSINESS

Your workflow could be the next case study.

You can explore AI without assuming your information must leave your environment. I can help assess whether a local-model architecture fits the task, where explicit rules belong and how your team should review the result.

Describe the material you need to analyze, the decision being supported and your infrastructure restrictions. We can scope a local workflow around those constraints.

Work directly with Itzhak Shaul, founder of Methodspire and Principal Software Architect with 30+ years of hands-on experience. Start with a focused conversation about the workflow, constraints and delivery scope.

Selected freelance experience of Itzhak Shaul. Project details are summarized to protect client confidentiality.