Inspiration
Self-hosted applications rarely fail in one obvious place. A process may be missing, a port closed, an HTTP endpoint unhealthy, an installed version different from its contract, or DEV inventory inconsistent with the real PRD runtime. Operators normally correlate all of those signals manually.
What it does
Visor AI adds a read-only AI operations layer to the existing Visor Apps platform. It collects a bounded, sanitized snapshot of application state, manifests, installations, processes, ports, local HTTP checks, and optional logs. GPT-5.6 turns that evidence into an executive summary, prioritized incidents, probable causes, safe recommended actions, and explicit missing data.
The operator can analyze the whole environment or focus on one app and its OAuth proxy. A local baseline enables before/after incident analysis. The model never starts, stops, restarts, or repairs a service: operational actions remain separate and require human confirmation.
How we built it
Codex was the primary development environment for repository inspection, architecture tracing, implementation, debugging, bilingual Streamlit UI work, tests, and documentation. GPT-5.6 runs through the OpenAI Responses API and reasons over the sanitized operational snapshot. The UI records the model, Response ID, and token usage for traceability.
Visor Apps existed before Build Week. The new contest work is clearly isolated and documented: the AI diagnostics module and view, loopback-only HTTP health probes, Spanish/English output, single-app scope, persistent three-report history, Markdown export, deterministic baseline comparison, Portfolio integration, and focused tests that mock the API.
Challenges
The main challenge was distinguishing declared configuration from observed reality. A service can be configured as disabled while still running, and an open port does not prove that the application is healthy. We normalized those signals before asking the model to reason and required it to separate evidence from hypotheses.
We also kept the feature safe and inexpensive: logs are opt-in, secrets are recursively redacted, context size is capped, HTTP checks are restricted to loopback, and the AI has no state-changing tools.
Accomplishments
- Turned a mature operations dashboard into an explainable AI operations assistant.
- Diagnosed real runtime/configuration drift that simple dashboards misreported.
- Added a convincing before/after incident workflow while keeping API usage low.
- Preserved existing behavior and added 21 passing automated tests.
What we learned
Operational AI is most useful when it receives normalized evidence rather than raw logs. Explicit uncertainty and verification steps make the output more trustworthy than an unconstrained operations chatbot.
What's next
We plan to add domain-aware HTTPS checks, app-specific log excerpts, previewable remediation plans with explicit approval, and pluggable inventory providers for more operating systems.
Log in or sign up for Devpost to join the conversation.