Hello Again

Inspiration

Meeting someone for the second time can be more stressful than meeting them for the first. A name disappears, and a profile photo may be years old, carefully posed, or simply look nothing like the person standing in front of you. A promising conversation stalls—not because you do not care, but because face and name memory are hard.

Hello Again starts from a simple belief: we should stop treating face memory as a test people must pass in order to belong.

What it does

Hello Again is a smartphone-only reunion assistant. Two people exchange profiles once using a QR code, the phone camera, or a six-character code. At a future event, each person can voluntarily register one location snapshot, valid for one hour. When an exchanged friend is within approximately 150 meters, Hello Again restores their name, organization, the user's original private note, and a useful conversation cue.

The location action takes one snapshot. It does not continuously track movement or update in the background. Friends see only a nearby result, never the other person's exact coordinates.

Users can save, edit, and delete a private original note and nickname. Phone-keyboard dictation works in real time without a separate speech service. A private caution flag is available, but it is set and removed only by the user. AI never decides or recommends whether a person is dangerous, trustworthy, desirable, or a “caution person.”

Because profile photos can be old or unhelpful, a user can ask Hello Again to create a face view and a full-body, photorealistic fictional memory image from written visual traits. No real photograph is used as the portrait-generation input. The output is clearly labeled as AI-imagined, may be inaccurate, and is not for identification or reconstruction of the real person.

Consent and privacy by design

Before login, guest mode, or the judge demo, Hello Again requires three separate, explicit confirmations:

• agreement to the Terms of Use;
• agreement to the Privacy Policy and personal-data handling; and
• consent to send written appearance notes to OpenAI for photorealistic fictional memory-image generation, together with confirmation that another person's information will be entered only with lawful authority or permission.

Consent is versioned and stored with the profile. Existing profiles must accept the current version before using AI features. Settings includes permanent deletion of the profile, exchange history, private notes, and saved portraits.

Hello Again does not perform facial recognition, face matching, biometric enrollment, or identity scoring. An optional user-supplied profile avatar is handled as ordinary profile data; it is not used for recognition or as input to the imagined-portrait generator.

How I built it

This is an individual submission created by Fumiya, working with Codex. An early planning document contained a three-person team template, but the submitted product was actually built by one human entrant plus Codex.

The product is a mobile-first React and TypeScript application running as a Cloudflare Worker through vinext. Cloudflare D1 stores profiles, versioned consent, exchanges, private notes, one-hour location snapshots, and portrait choices. Cloudflare R2 privately stores selected generated images. QR generation and phone-camera scanning make exchange installation-free. A one-tap judge mode seeds twenty fictional English contacts, ten of them near Shibuya Solasta Conference.

The interface supports Japanese, English, Simplified Chinese, Korean, Spanish, French, German, and Portuguese. Terms of Use and the Privacy Policy are published in English and Japanese.

How OpenAI is used

GPT-5.6 is integrated through the Responses API with strict Structured Outputs and store: false. It extracts only explicitly stated facts, topics, and visual traits and creates concise reunion prompts. The system instruction prohibits inventing sensitive attributes and prohibits evaluating a person's danger, trustworthiness, personality, desirability, or social value.

GPT Image 2 creates a face close-up and a head-to-toe image from written visual details. Photorealism is intentional: this is a practical memory aid. Prompt constraints preserve explicitly written gender presentation, age range, body build, clothing, hair, and accessories without beautifying away important traits. The prompt also forbids identity reconstruction, public-figure resemblance, unstated sensitive traits, logos, and watermarks.

A deterministic fallback keeps the core judge path available if an external AI call is temporarily unavailable.

How I used Codex

Codex was my development partner throughout the primary Build Week thread, whose Session ID is included in the submission.

Codex helped me:

• turn the original idea into a deployable smartphone-only architecture;
• research browser Bluetooth and background-location limitations;
• implement the responsive interface, eight-language system, QR and camera flows, verified-email restoration, APIs, D1 migrations, R2 portrait storage, and OpenAI integrations;
• reproduce and fix phone and desktop UI regressions;
• build twenty fictional English judge personas and a deterministic demo path;
• add safety and privacy regression tests;
• create, validate, and publish the English demo video; and
• deploy the working public application.

I made the consequential product, privacy, UX, and business decisions as the sole human entrant. I rejected unreliable browser Bluetooth, chose explicit one-hour location snapshots, kept notes and caution flags private, prohibited AI person-risk classification, and required explicit consent for AI image processing.

Challenges

The hardest challenge was preserving the magical moment of knowing a friend is nearby without dedicated hardware or pretending that mobile browsers can provide dependable cross-platform background Bluetooth. The constraint led to a clearer design: one location snapshot, requested by the user, valid for one hour, and never exposed as exact coordinates to another person.

The second challenge was image fidelity without identity reconstruction. Written traits such as gender presentation, age range, clothing, hairstyle, and body build must remain visible for the image to be useful. The product therefore generates both face and full-body views, uses strong fidelity constraints, and labels every result as fictional and non-identifying.

The third challenge was drawing a firm line between memory assistance and automated judgment. GPT structures explicit notes, but the model schema contains no caution-person recommendation. Any private caution flag is entirely the user's manual action.

Accomplishments

• A working public judge demo with twenty memorable fictional English contacts
• A coherent smartphone experience requiring no external hardware
• Explicit, versioned first-use consent for Terms, Privacy, and AI image processing
• GPT-5.6 Structured Outputs without AI person-risk classification
• GPT Image 2 face and full-body fictional memory images
• Verified-email restoration, QR/camera exchange, multilingual UI, and private note editing
• One-hour proximity matching without continuous tracking or facial recognition
• Account and stored-data deletion from Settings
• Ten passing regression tests and a reproducible production build

What I learned

Platform constraints can improve a product. Losing background Web Bluetooth forced me to design a clearer consent model and a demo that anyone can open immediately. I also learned that AI memory support is most trustworthy when the original human note remains editable and visible, generated structure stays secondary, and human judgment is never delegated to the model.

What's next

A future native iOS and Android version can combine coarse location with Bluetooth Low Energy and opt-in notifications. It would broadcast rotating anonymous identifiers rather than names or fixed IDs, use no dedicated beacon, and resolve proximity only when both people have already exchanged and enabled event mode.

An Even G2 companion could later display the same name, private context, and conversation prompt on the glasses HUD while the smartphone remains the consent, account, networking, and location hub. Conference, professional-community, alumni, and accessibility-focused workplace partnerships are the initial business path.

Built With

• codex