SellPilot

Inspiration

Selling a used item looks simple until the seller has to identify it, describe its condition, choose a price, write a trustworthy listing, and answer buyers. Existing AI tools can generate copy, but they often hide uncertainty and make confident claims without showing evidence.

What it does

SellPilot turns one to four product photos plus optional seller context into a structured selling guide. It separates seller-provided facts, AI detections, and uncertain information. The result includes probable product identification, visible condition, missing details, a recommended next photo, three clearly labeled price scenarios, listing copy, and quick buyer responses.

SellPilot does not claim to read live marketplace data, verify authenticity, or publish automatically. When OpenAI is unavailable, it offers a simulation that is explicitly labeled and states that it did not inspect the photos.

How we built it

The application uses Next.js, React, and TypeScript. A server-side endpoint validates and limits each request before calling the OpenAI Responses API with GPT-5.6. Images are sent with high detail and the model returns strict Structured Outputs validated against a JSON schema. Results are displayed with provenance, evidence, confidence, and uncertainty. Inventory and drafts remain in the current browser for this Build Week version.

How GPT-5.6 is used

GPT-5.6 performs the multimodal reasoning: it reconciles several photos with human context, identifies visible evidence, expresses confidence, and produces a broad structured response. The integration uses low reasoning effort, high image detail, strict JSON output, and store: false.

How Codex helped

Codex acted as product collaborator, engineer, QA lead, and release manager. It audited the prototype, separated real and simulated functionality, designed the structured schema, integrated GPT-5.6 server-side, added validation and safety controls, tested responsive behavior and accessibility, fixed dependency risks, prepared documentation, and deployed the public demo.

Challenges we ran into

The hardest challenge was making AI uncertainty useful instead of hiding it. We turned missing evidence into a next action: tell the seller exactly which photo or fact would improve confidence. We also built a resilient fallback without presenting simulated output as real AI analysis.

Accomplishments that we're proud of

SellPilot is transparent about provenance and limitations, works without an account in its public demo, includes a real server-side OpenAI path, provides reviewable sales content, and passed its automated tests, TypeScript checks, linting, and production builds.

What we learned

Trust improves when the product shows what the model saw, what came from the user, what remains uncertain, and what the seller should verify next.

What's next for SellPilot

The roadmap includes accounts, durable storage, multi-company workspaces, distributed limits, backups, consent controls, team roles, audit logs, and marketplace integrations.

Built With

• gpt-5.6