plateful

Inspiration

Going out should feel simple and safe. Friends with allergies often spend more time reading menus than enjoying the meal. We wanted a tool that understands plain language like late night spicy noodles near me and returns great options with clear allergen signals backed by real evidence.

What it does

Plateful recommends nearby restaurants that fit your cravings and your safety needs. Type a prompt or grab a quick top ten near you. For each place, Plateful shows rating, distance, and an allergen summary. It finds text menus or menu photos, runs OCR, and estimates allergen probabilities by dish with short evidence snippets and links.

How we built it

Frontend NextJS, React, Tailwind for fast UI and optimistic updates
Backend Python Flask APIs for search, allergen analysis, images, and review stats
Agents Letta for stateful orchestration and tool selection, Claude Haiku 4.5 for structured reasoning and extraction
Data BrightData for Google Maps business info and reviews, web scraping for menus and venue images
OCR Tesseract or GCV depending on source quality
Ranking Distance, rating, review count, prompt fit, and allergen fit with tunable weights
Infra Redis cache for geo results and reports, Postgres for entities and provenance, structured logs and traces for observability

Challenges we ran into

Finding reliable menu sources and respecting robots rules
OCR noise from low resolution or stylized menu images
Normalizing cuisines and dish names across languages
Calibrating allergen probabilities without overstating confidence
Keeping latency low while running multiple tools per query
De-duplicating and filtering venue images for relevance and safety

Accomplishments that we’re proud of

End to end pipeline from natural language prompt to ranked results with evidence
Dish level allergen summaries that cite menu text or OCR extracts
Stable tool schemas with idempotent jobs and caching for fast nearby queries
Clear UX with interpretation chips, confidence banners, and provenance links
Solid test coverage for API contracts and prompt outputs

What we learned

Users want speed and trust. Evidence and confidence scores matter as much as rankings.
Geo caching by geohash and sensible defaults dramatically improves perceived performance.
Structured JSON outputs from the model reduce parsing errors and make retries predictable.
Cuisine priors help, but must stay as hints. Evidence always wins.
Small quality of life details like distance in meters and simple allergen badges reduce friction.