Inspiration

For an independent consultant, one inbox can hold a launch blocker, a request for next week, a dependency on a client, and a newsletter. The practical problem is deciding what deserves attention while keeping the original context close enough to check.

What it does

Inbox Flight Deck turns sample or pasted messages into four lanes: Do now, Up next, Waiting, and Filed away. In live mode, a Gemini agent proposes a concrete next action and rationale for each message. Every proposal includes an exact source quote that the server verifies against the original message body.

The person reviews the plan, changes categories if needed, and selects which decisions to apply. Those decisions move to a local action board; unselected messages stay in the inbox. The board survives a reload, exports a Markdown brief with evidence, and supports undoing the last batch. The proof of concept never sends replies or deletes mail.

How it was built

The project started in a new empty folder during the submission window. Codex was the coding agent. The Devpost Learn Skill Pack was installed and used to structure scope, product requirements, technical specification, implementation slices, verification and shipping. Planning documents were committed before application code. The owner delegated concept selection and implementation; workflow adaptations are disclosed in the repository.

The stack is TypeScript, Node.js, Express, React, Vite and Zod. A server-side Gemini call returns structured decisions. Validation rejects invented or duplicate message IDs, missing decisions and quotes that do not occur in the source. A separate deterministic rehearsal mode is visibly labeled and works without credentials. Source and fictional sample data were created for this project; standard dependencies and the required Skill Pack are disclosed.

Challenges

One real integration issue was model availability: the initial Gemini 2.5 endpoint rejected new-user access and recommended Gemini 3.8 Flash. The model was changed and checked with a real API call before handoff. Another design constraint was separating an AI proposal from execution: selecting and applying a batch is a distinct human action, and the prior state remains available for undo.

Accomplishments

The implemented loop runs from inbox to validated plan to selected board changes, export and undo. Verification covers strict typechecking, planner/API/transaction tests, a production Chrome journey, mobile layout and a real Gemini-backed browser journey. Exact results are documented in the repository rather than presented as performance or user-study claims.

What was learned

The build exposed two useful distinctions. An exact source quote can be verified mechanically, while a priority decision still needs human judgment. A valid proposal also needs a separate, reversible execution step. Turning those distinctions into acceptance criteria made it possible to test that unselected messages stay untouched and that undo restores the prior state. The provider integration also showed why available model access must be checked with a real request rather than inferred from a model listing. These are project findings, not a claim about any participant’s personal learning.

What's next

Possible extensions are a read-only mailbox connector, better handling of relative deadlines, and richer review controls. They are not part of the current proof of concept.

Testing instructions

Requires Node.js 22+. From the public repository: npm ci, npm run build, npm start; open http://localhost:3000. Rehearsal works without credentials. Make a plan, inspect a quote, uncheck one proposal, apply seven decisions, export, reload, and undo. Configure server-side GEMINI_API_KEY in .env and restart to enable Live Gemini. Never paste a key into Devpost. The recorded video must show the working project because reviewers may judge without running the repository.

Attribution / AI disclosure

Application code, fictional data and submission materials were prepared with Codex. The required Devpost Learn Skill Pack and standard npm dependencies are credited in the repository. The demo uses automated capture of the actual local application and a real Gemini request, with AI-assisted English captions. It contains no generated likeness, voice, music or stock footage. The owner's concept/implementation delegation and workflow adaptations are documented in docs/WORKFLOW.md; no owner hands-on review or personal learning is claimed.

Built With

Share this project:

Updates

Submission history