IncidentPilot

Inspiration

Small software teams and application-support operators often move between service status, logs, repair commands and incident notes. A restart can succeed while the user-facing feature remains broken. A familiar symptom can also have a different cause. The useful outcome is a verified recovery—or an honest handoff with evidence—rather than a confident command transcript.

What it does

Select an application, describe the symptom and start an investigation. IncidentPilot reads current evidence, selects bounded tools, checks permission outside the model, and verifies the actual service outcome after action. Four cards explain what broke, what was done, whether it recovered and what the user must do. Expand the evidence for technical detail and download the same-run report. Missing access or key evidence produces a human handoff.

The test build includes a real HTTP application and a real queued background worker owned by the project. They require no company infrastructure. A connector interface keeps the core independent of a specific ERP, CRM or service type.

How we built it

Python, Strands Agents, the existing IncidentPilot ExecutionGuard, a small connector contract, standard-library local services, and a dependency-free browser interface. The competition provider is NVIDIA nvidia/nemotron-3-super-120b-a12b on Nebius Token Factory. The model is responsible for interpreting information, selecting tools and adapting decisions; the program enforces budget, scope, permissions and actual recovery checks. There is no provider fallback.

What we tested

Offline service tests, isolated SDK tests, real-model runs and browser checks are reported separately in VALIDATION.md. A six-case comparison uses a fixed SOP, the same model in a generic agent, and IncidentPilot. Six CLI real-model cases: four recoveries and two permission/evidence handoffs. Two additional live browser investigations passed on 2026-09-29. Small evaluation results are not evidence of production reliability.

October 5 reliability update

The updated source (e16249b) repairs existing behavior: investigation tracking resumes after a lost submission response, mode labels follow the running server state, cancellation preserves reports and works during startup, stale stop responses do not affect later runs, and Linux runtime dependencies install correctly. Existing budget ledgers are validated and a running server cannot silently recreate a missing ledger. No incident capabilities or paid hosting were added.

Validation for this revision: 35 Python tests, 9 UI regression tests and 6 real-browser offline scenarios passed. The new checks made no paid model calls. They complement the earlier real-model evidence; they do not constitute new real-model validation.

Origin

This is a significant competition-period update to an existing project, not a new-from-scratch claim. See PROVENANCE.md in the repo. Existing September 9 documentation commits were retained; competition-period work added Nebius/Nemotron integration, connected-app investigation, independent recovery verification, evidence/reporting UI, and free judge hosting.

Nebius / NVIDIA feedback

Chosen integration: NVIDIA nvidia/nemotron-3-super-120b-a12b through Nebius Token Factory's OpenAI-compatible API, wrapped by Strands OpenAIModel. Official cookbooks provide exact regional endpoint and model id; function-calling docs cleanly separate model intent from application execution. Authenticated catalogue (2026-09-28) reports tools/reasoning support and rates $0.30/M input, $0.90/M output. Observed limitation: model initially skipped useful diagnostic logs on a locked task and handed off too early; investigation-prompt correction led to log-based adaptation and verified recovery. Suggested improvement: expose remaining promotional credit and exact expiration via a documented read-only API so an agent's local ledger can match account balance for judge-window access.

Live access

Nebius $25 Token Factory promo redeemed 2026-09-30; balance $25; provider email stated credits do not expire once applied. Live access intended through judging end 2026-12-15 via free Render URL. No top-up / paid services used for this submission.

Free demo storage limitation

The demo remains on Render's free instance. Host restarts or redeployments can discard local reports and the local usage ledger, so download reports that need to be retained. The local ledger is not a durable account-wide spending cap or a provider-balance check. Persistence across replacement host instances remains a limitation.

Built With

• compatible
• mit
• nebius
• nemotron
• nvidia
• openai
• python
• strands
• strands-social-recommender