NeuroRoute

Inspiration

I kept noticing the same gap every time we looked at navigation apps: they're built entirely around speed. Fastest route, shortest time, fewest turns, never once asking whether a person can actually handle that route. For autistic travelers, that gap is huge. A path that looks perfectly short on Google Maps can run straight through a metro exit with sudden alarms, a market footpath that turns into a crowd surge, or a wall of flickering LED signage. None of that shows up on a map, yet it's often the real reason a trip gets cancelled, or ends in sensory overload before the destination is even reached. We wanted to build something that treats sensory exposure as seriously as travel time, because for a lot of people, it's the bigger barrier.

What it does

NeuroRoute is a sensory-aware navigation tool for autistic travelers and their caregivers. A user enters an origin, a destination, and a sensory profile: whether they want to avoid sudden loud noise, dense unpredictable crowds, or harsh or flickering light. NeuroRoute then: Scans the route corridor for likely sensory hazards using an AI hazard-mapping layer Generates two paths: the standard route and a calmer, sensory-optimized alternative Renders both routes live on an interactive map, with hazard zones marked and labeled Produces a tactical prep plan, concrete step-by-step guidance a caregiver can actually use, like putting on noise-reducing headphones before a metro exit or taking a wider, tree-lined pavement instead

How we built it

The frontend is built in Streamlit, with a sidebar for trip parameters and sensory filters and a live dashboard for results. The hazard-detection layer calls the Gemini API, prompting it to reason geospatially about a given route and return a structured JSON object with coordinates for the standard path, the alternative path, and flagged hazard zones. I render everything with Pydeck (deck.gl), using PathLayer for the two routes and a ScatterplotLayer for hazard markers, on a free Carto basemap so the map works without needing a paid token. Because hackathon demos can't afford to break mid-pitch, I built a simulated-data fallback: if the AI call fails or no API key is provided, the app gracefully falls back to realistic pre-computed data for the same route, so the experience never crashes in front of judges.

Challenges we ran into

The biggest challenge was getting the AI's output into a shape I could actually render. Gemini doesn't always return clean JSON on the first try, so I had to be strict about prompt formatting and add parsing safeguards for stray markdown or malformed responses. I also initially used a Mapbox-style map URL that silently failed without an API token, which meant the map was rendering blank. That took some real debugging because it failed quietly instead of throwing an obvious error. The deeper challenge, though, was translating "sensory accessibility" into something concrete and locatable. It's easy to say "avoid crowds." It's harder to define what that means at a specific metro exit or market footpath in a way that's actually useful and not just generic.

Accomplishments that we're proud of

I'm proud that this isn't just a concept slide. It's a working, demoable app with a real AI layer, a real interactive map, and a fallback architecture that means it won't break during a live demo. I'm also proud of grounding this in a real-world location (Connaught Place, New Delhi) rather than a generic placeholder city, and tying the project's purpose directly to organizations like the Akhil Autism Foundation, which already does this kind of support work indoors. NeuroRoute is my attempt to extend that same care outdoors, into the actual streets and stations people have to navigate.

What we learned

I learned how much nuance sits inside "accessibility" as a design problem. It's not one slider, it's at least three independent dimensions (noise, crowd density, light) that all interact differently depending on the place. I also learned a lot about working with LLMs as a structured-data layer rather than a chat interface: getting an AI model to reliably return usable, well-formed geospatial JSON took more careful prompt design than I expected. And on the product side, I learned that resilience, making sure a demo degrades gracefully instead of crashing, matters just as much as the headline feature.

What's next for NeuroRoute

Right now, hazard detection is AI-inferred rather than measured. The next step is grounding it in real data: live transit crowd APIs, municipal noise-complaint datasets, and open city data, instead of relying solely on AI inference. I'd also like to build a caregiver feedback loop, so that if a flagged hazard turns out to be inaccurate, or a real one was missed, the system gets better at scoring that route over time. Longer term, I want to pilot NeuroRoute with autism therapy centers and special schools in the NCR region, partner with organizations like the Akhil Autism Foundation, and move from a Streamlit prototype to a mobile app that someone could actually use mid-trip, not just in a demo.