OnThePorch

The Public Safety and Community Sentiment project aims to help Dorchester residents and community leaders access timely, trustworthy information about local public safety issues and community events by combining official city data with verified community reporting. During Fall 2025, the student team delivered a working AI-powered community assistant that integrates a chat-based interface with a filterable event calendar. Users can ask natural language questions such as “What’s happening this week?” or browse upcoming events directly from a sidebar. The system currently ingests data from Dorchester Reporter, Boston.gov APIs, community newsletters, and uploaded documents, with automated daily updates. A key focus of Fall 2025 was improving data ingestion, source coverage, and usability rather than expanding analytical complexity. The team implemented automated pipelines for newsletters and public data, ensured clear source attribution with links to original reporting, and designed a UI that was approved by the client for clarity and accessibility. Client feedback emphasized the importance of avoiding AI editorializing, particularly when discussing sensitive public safety incidents. The preferred interaction model is to present verified facts first, then optionally offer historical trends or sentiment context only when explicitly requested. This insight reframed the system’s role as an assistive information tool, not an interpretive or opinionated one. While the current system functions as a proof of concept, several known limitations remain, including incomplete shooting incident location data, gaps in recent datasets, and the need for structured user testing with community members. This semester’s continuation project will focus on strengthening reliability and trustworthiness, expanding newsletter ingestion, and developing a clearer framework for separating factual reporting from trends and optional context. The intent is to conclude the student-driven development cycle with a more robust prototype that can be evaluated through user testing and positioned for potential productization or handoff to professional development resources. In addition to newsletter ingestion, this semester will extend the platform to support community conversation audio ingestion. The team will prototype a lightweight backend/admin upload interface where a trusted partner can submit an audio file and required metadata (meeting date/time, location, context, agenda, and any official roles present). The system will then generate a transcript, redact personally identifiable information for community participants, and index the content so that the assistant can answer questions by quoting the transcript explicitly (for example, “A community member said…” with a timestamp/date) rather than mimicking a community voice. This work supports the project’s core trust requirement: the assistant should provide verifiable information with clear sourcing, and only use transcript-based sentiment or context when the user explicitly requests it. (Please confirm with client before continuing this part) Project recommendations & next steps: (Ideal Output & Final Deliverables) Recommended Next Steps • Reintegrate and validate shooting and shot-spotter data with accurate latitude/longitude information. • Implement a prompting framework that clearly separates: • factual reporting • historical trends • optional sentiment or contextual analysis • Expand automated ingestion to additional community newsletters identified by the client. • Conduct usability testing with at least 20 community members to assess trust, clarity, and usefulness. • Add power-user functionality such as CSV exports and structured summaries for reports. • Building a mechanism to ingest/upload new community conversation audio files, similar to how newsletters are ingested. • Capturing required metadata at upload time (date/time, place, meeting context, agenda if available, whether public officials were present, etc.). Things to Avoid • Do not allow the AI to automatically editorialize or blend historical quotes with current incidents. • Do not surface internal spreadsheets, raw transcripts, or documents containing names. • Avoid overloading responses with sentiment when users are asking for simple factual information. Common Blockers • Incomplete or delayed public datasets (especially for recent years). • Sensitivity around how incidents and quotes are framed. • Token and context limitations when working with large text corpora.