Inspiration
Quality inspectors often work with long SOPs, incomplete visual evidence, and decisions that are difficult to audit. A conventional AI assistant may produce a confident answer without proving that the available evidence supports it.
Inspection Copilot explores a safer approach: every inspection must provide structured evidence, connect its findings to the applicable SOP, and fail closed to human review whenever the evidence or model output is insufficient.
What it does
Inspection Copilot turns a product image and a structured inspection procedure into an auditable result:
- pass, fail, or needs human review;
- structured observations connected to SOP requirements;
- explicit confidence and evidence-completeness checks;
- fail-closed handling of timeouts, refusals, rate limits, and invalid output;
- human override with a required rationale;
- sanitized provenance and reproducible evaluation records.
The demonstration uses repository-owned synthetic cases rather than private production data.
How we built it
We designed and implemented the project in OpenAI Codex using a test-driven, incremental workflow.
The application is written in Python with Pydantic contracts and a Streamlit interface. GPT-5.6 is integrated through the OpenAI Responses API behind a provider-neutral inspection interface.
GPT-5.6 produces a structured visual assessment, while deterministic policy logic decides whether the evidence supports an automatic verdict or requires human
Log in or sign up for Devpost to join the conversation.