# Project export: Public Transit AI This document was generated by HackStack to give an AI agent context about a hackathon project. Sections are labeled with their provenance; content marked as truncated was cut to keep this document small. ## Project metadata - Hackathon: UC Berkeley AI Hackathon 2025 - Tagline: We turn raw, messy transit complaints into fast, action-ready charts with Claude-written analysis, revealing concert trash surges, rush-hour delays, and harassment hotspots - no data analysts needed. - Devpost: https://devpost.com/software/transit-ai-nro2p8 - GitHub: not linked - Team: contributor stats unavailable ## Devpost submission (written by the team) ### Inspiration Transit agencies drown in citizen complaints but lack staff to sift patterns fast enough to act. We wanted an AI co-pilot that surfaces the "why" behind the numbers -- not just another dashboard. At the same time, we believe in transparency: riders should be able to make their complaints public so everyone can stay informed, empowered, and safe together. ### What it does Ingests complaint records - we created a synthetic mirror of SFMTA public feedback data Slices data automatically into actionable segments: rush hour vs. off-peak, equipment failures in heat waves, concert/sports event days, high-incident stops, and more Plots each slice with Chart.js for intuitive exploration Summarizes the story using Claude 4 Opus. Our prompt includes examples and special logic (e.g., merging adjacent stations so that insights move beyond single stops) to generate clear, high quality insights ### How we built it ✨ Synthetic data: Generated 1,000 realistic complaints mirroring public transit patterns across five years. πŸ›’οΈ Supabase Postgres stores incidents; Claude adds auto-tags to help organize them. 🧠 Claude 3 Opus summarizes each chart with markdown insights. Prompt includes few-shot examples, adjacent-station logic, and chart JSON. πŸ“ˆ Next.js 13 + React-ChartJS-2 render the dashboard. Plugins add baselines, hotspots, and annotations. 🧊 Caching with Supabase: Summaries are stored under chart_id so we don’t hammer Claude every time. ### Challenges we ran into No public datasets or PII allowed: We built a synthetic dataset statistically resembling real complaints while preserving privacy LLM anchoring bias: Claude hallucinated "no reports" when the JSON started with multiple zeroes (midnight-4 AM where no reports are made). Solving via pre-filtering and prompt tuning to prevent early-token bias from dominating. ### Accomplishments we're proud of 🌍 Aligned with UN SDGs: SDG 11 (Sustainable Cities & Communities): boosts trust and usage of public transit SDG 5 (Gender Equality): safer commutes through proactive harassment alerts SDG 3 (Good Health & Well-being): crowding insights help riders avoid unhealthy over-packed vehicles SDG 9 (Industry, Innovation & Infrastructure): AI-driven, low-cost analytics for civic infrastructure. SDG 11 (Sustainable Cities & Communities): boosts trust and usage of public transit SDG 5 (Gender Equality): safer commutes through proactive harassment alerts SDG 3 (Good Health & Well-being): crowding insights help riders avoid unhealthy over-packed vehicles SDG 9 (Industry, Innovation & Infrastructure): AI-driven, low-cost analytics for civic infrastructure. ⚑ From CSV β†’ dashboard β†’ AI summary in ~3 s average round-trip πŸ’Έ Entire stack runs on Supabase free tier plus < $5/month for Claude API calls ### What we learned Synthetic data can be just as powerful for prototyping real-world civic tech Few-shot prompting (feeding 5-7 formatted examples) gave Claude a consistent executive tone; zero-shot was hit-or-miss. Don’t feed your LLM raw zeros without thinking about positional anchoring bias. ### What's next for Transit AI πŸ“± Public-facing dashboard for real-time safety & crowding info - riders can choose to make complaints public or anonymous. Women and vulnerable riders get real-time flags for trains/stations with rising harassment reports from others (crowd-effort). Live crowding + peak-hour heat-map, letting anyone plan a quieter or less-packed trip. πŸ—“οΈ Event integration (Eventbrite, city APIs) to detect cleanliness/crowding issues around big venues. πŸ”„ Swap in real 311 data from public transportation agencies - exact schema, just a simple Supabase import. Capture more insights. πŸ“© Share weekly PDF insight packets with the city’s transit ops team. πŸ’¬ Community feedback loop. Open source the API so advocacy groups can build add-ons (e.g., accessibility widgets, language-specific safety tips). 🧠 RAG-based querying (Retrieval-Augmented Generation) - use Supabase pgvector to push raw reports into a vector store and answer questions like, "Why are failures high at 85 Β°F?" by doing a similarity search for relevant historical incidents πŸ› οΈ Trip-planner plugin β€” let journey apps like Google Maps pull in safety & crowding scores. πŸ“Š ML classifiers β€” small models (logistic regression, XGBoost) to auto-detect likely harassment reports or maintenance issues. ## README (from the GitHub repository) No README available. ## Detected evidence (automated analysis) No repository was indexed for this project. Claimed technologies below could not be checked against code. - Anthropic (technology) β€” claimed on Devpost, not found in the code - Next.js (technology) β€” claimed on Devpost, not found in the code - Node.js (technology) β€” claimed on Devpost, not found in the code - OpenAI (technology) β€” claimed on Devpost, not found in the code - React (technology) β€” claimed on Devpost, not found in the code - Supabase (technology) β€” claimed on Devpost, not found in the code - Tailwind CSS (technology) β€” claimed on Devpost, not found in the code - Vercel (technology) β€” claimed on Devpost, not found in the code ## Codebase structure No repository index available. ## Key source files No repository index available; no source files included.