A public case study for an agentic verification workflow that connects systems engineering, model-based architecture review, and retrieval-augmented evaluation.
This repository intentionally contains no production source code, datasets, copied standards material, grant material, cloud deployment files, credentials, or build logs. It is a sanitized portfolio artifact that explains the engineering approach at a public boundary.
- Agentic review of systems-engineering artifacts
- RAG-grounded evidence selection for technical review
- Traceability and faithfulness checks across architecture inputs
- A public-safe architecture pattern rather than a source dump
Complex cyber-physical systems produce requirements, architecture models, design assumptions, and verification evidence across multiple tools. Reviewing those artifacts manually is slow, error-prone, and difficult to scale.
The project explored how an agentic verification workflow can help:
- parse structured architecture inputs
- retrieve relevant requirement context
- evaluate answer faithfulness against source material
- surface traceability gaps
- support human review instead of replacing it
The demonstrated pattern used modular layers:
- Ingestion layer for source documents and structured model artifacts.
- Parsing layer for model-oriented text and requirement-like records.
- Retrieval layer for contextual evidence selection.
- Verification layer for faithfulness, precision, and recall-style scoring.
- Review UI for inspecting generated conclusions and evidence.
The public takeaway is the pattern, not a full implementation dump.
- Agentic decomposition of a systems-engineering review task.
- Retrieval-augmented evidence grounding.
- LLM-as-reviewer scoring for faithfulness-style checks.
- A web-facing mission-control style interface.
- Batch-style processing of technical source material.
The primary user is a technical reviewer, systems engineer, or portfolio reviewer who needs to understand whether the demonstrated verifier pattern keeps evidence visible while preserving a public release boundary. The job-to-be-done is to assess the workflow shape without requiring access to private source material. A success signal is that the reviewer can explain the problem, architecture pattern, human-review role, and excluded materials from the public docs alone.
This repo is documentation-only. The quickstart for local review is to read README.md, docs/CASE_STUDY.md, and docs/RELEASE_BOUNDARY.md, then run the pre-release verification command documented in VERIFICATION_PLAN.md before publishing changes.
- Original source code.
- Standards, vendor, government, or proprietary PDF material.
- Grant application content.
- Cloud deployment configuration.
- Planning notes, build logs, or operating instructions.
- API keys, model-provider configuration, or credentials.
- Robotics glossary and SysML examples: https://github.com/camirian/robotics-ontology-public
- Sim-to-real control examples: https://github.com/camirian/sim-to-real-control-systems-public
- Articulated manipulation examples: https://github.com/camirian/articulated-robot-manipulation-public
This case study is published for portfolio review. See LICENSE.