IllustrationLearning a way across.
Comparing a Q-table and a deep Q-network in the FrozenLake environment.
Explore the projectAI / Software architecture
AI with a paper trail.
A local multi-agent system that makes evidence, review and execution traces part of the workflow.
01 / The challenge
The question behind the system.
A generated answer is only part of a useful document-analysis system. The application also needs to record which evidence was used, check claims and stop when it cannot validate a result.
Secure Agent Governance Engine coordinates researcher, analyst and reviewer agents around local documents. Ollama provides local inference; a FastAPI application exposes the workflow and its traces. Explicit policies, source checks and bounded correction attempts make the route to an answer inspectable.
02 / The approach
From architecture to implementation.
- 01
Ingest local text, Markdown, PDF and DOCX files, retrieve relevant chunks and keep evidence identifiers attached to them.
- 02
Separate research, analysis and review responsibilities, with deterministic checks alongside model-based review.
- 03
Declare agent permissions and correction limits in YAML policies, and record workflow attempts in SQLite and JSONL.
- 04
Expose the workflow through a local web interface and FastAPI, with automated tests and a Docker setup.
03 / Outcomes
What the work shows.
- A local application connects document retrieval, analysis, review and bounded correction in one observable workflow.
- The repository provides test coverage for orchestration, document ingestion, validation, traces and API behaviour.
04 / Scope & next steps
Where the work stands.
- The application is a local demonstration. Its API has no authentication or user management and is not presented as a production service.
- Prompt-injection checks and semantic review can miss hostile content or incorrect claims; they do not guarantee safe or correct model output.
Continue exploring
