Inspiration
What it does
How we built it
Inspiration
Many people want to go out but get stuck at the first question: “What should I do today?” Traditional travel and local-discovery apps often make this harder by returning long lists of attractions, rankings, and filters.
We wanted to build a small AI concierge that understands one honest sentence about the user's current mood, investigates what is actually available nearby today, and makes one practical decision instead of presenting another list.
What it does
Nearby Now turns a short mood statement into one nearby, low-cost activity that the user can begin today.
The agent:
- understands the user's current energy, social preference, and willingness to travel;
- investigates both same-day public events and a no-event neighborhood roaming option;
- checks the date, time, location, availability, cost, and public sources;
- compares the candidates and recommends only one;
- provides a clear first step, three small actions, estimated time and cost, practical reminders, and supporting sources.
The interface supports instant Chinese and English switching, including the language of the AI-generated plan. Photos and music preferences are optional weak aesthetic signals and are never used for medical diagnosis or sensitive-personality inference.
If live information cannot be verified, Nearby Now does not invent an event. It clearly switches to an offline roaming plan and asks the user to confirm weather, opening hours, and local conditions before leaving.
How we built it
We built Nearby Now as a bilingual, installable PWA with a zero-dependency Node.js service.
GPT-5.6 is accessed through the OpenAI Responses API. The agent uses web search to investigate current public information, constructs both an event candidate and a neighborhood-roaming candidate, applies evidence gates, and then returns one structured recommendation.
Codex helped us design and implement the responsive interface, local agent service, bilingual experience, privacy boundaries, evidence-validation logic, offline fallback, smoke tests, documentation, and submission materials.
We also packaged the decision workflow as a reusable Codex skill called local-life-concierge, with its own recommendation contract, evidence requirements, privacy rules, and structured output format.
Challenges we ran into
The hardest challenge was making a recommendation that was both personal and trustworthy. Same-day event information can be incomplete, outdated, or missing essential details. We therefore required every event candidate to include a valid local date, place, usable time window, and public source before it could be recommended.
Another challenge was graceful failure. The application still needed to be useful when an API, model, search provider, location service, or public website was unavailable. We solved this with an explicit offline roaming mode instead of pretending that live verification had succeeded.
We also had to keep the Chinese and English interfaces, AI output language, structured response contract, mobile layout, and privacy behavior consistent across the entire application.
Accomplishments that we're proud of
We are proud that Nearby Now produces one bounded, immediately actionable decision rather than another overwhelming list.
We also built:
- evidence gates for time-sensitive public events;
- transparent offline fallback behavior;
- local-first photo color extraction and explicit upload consent;
- local API-key storage with masked configuration;
- an installable bilingual PWA for desktop and mobile;
- support for OpenAI and additional model providers;
- a reusable Codex skill that can operate independently from the web interface;
- smoke tests covering the interface, provider paths, search flow, and recommendation contract.
The project is open source under the MIT License and can run locally without an API key in offline mode.
What we learned
We learned that a useful local AI agent needs more than search and fluent text. It needs a clear decision contract, evidence thresholds, privacy boundaries, and permission to abstain when reality cannot be verified.
We also learned that personalization can remain lightweight. A short mood statement and a few optional aesthetic signals are often enough to reduce decision friction without constructing a permanent or invasive profile.
Most importantly, the best response is not always the most elaborate plan. It is the plan with the smallest realistic first step.
What's next for Nearby Now
Next, we plan to add regional adapters for weather, maps, public transport, opening hours, and official event sources. We also want to improve accessibility, build a formal recommendation-evaluation dataset, expand multilingual support, and add privacy-preserving preference synchronization across devices.
Our long-term goal is to make Nearby Now a trusted local action layer: one sentence in, one realistic plan out.
Log in or sign up for Devpost to join the conversation.