Inspiration
New York has systems that tell us what is happening, but what about what happens next? We built NYC Pulse to investigate how a single city disruption could create ripple effects across transportation, infrastructure, and public services.
What it does
NYC Pulse is an AI-powered urban intelligence agent. Users describe an NYC disruption, and the agent investigates potential second- and third-order effects using real NYC transit and 311 data.
It generates hypotheses, retrieves historical signals, checks the evidence, and produces confidence-aware assessments, monitoring recommendations, and suggested actions.
How we built it
We combined Mistral Large for reasoning, hypothesis generation, and tool selection with Elasticsearch Serverless for evidence retrieval.
Our Elasticsearch implementation includes:
- Geographic search over 1,488 MTA subway-stop records
- BM25 search and aggregations over 20,000 NYC 311 requests
- Semantic retrieval over 2,000 additional indexed representations of 311 records using Mistral Embed and Elasticsearch's native inference API
Our key innovation is the Evidence Loop: Mistral forms hypotheses, Elasticsearch searches for relevant evidence, and Mistral evaluates the results before producing a final assessment.
The interface is built with Python and Streamlit.
Challenges we faced
The biggest challenge was preventing plausible AI-generated predictions from being mistaken for verified facts. We addressed this by separating hypotheses from retrieved evidence, highlighting limitations, and using an additional verification stage.
What we learned
Combining agentic reasoning with geographic, lexical, and semantic retrieval makes urban data more useful for exploring complex scenarios. We also learned the importance of distinguishing contextual historical signals from direct causal evidence.
Accomplishments
- Built an end-to-end agentic investigation workflow
- Indexed 21,488 NYC transit and 311 records
- Integrated Mistral embeddings with Elasticsearch semantic search
- Implemented dynamic hypothesis verification
- Successfully tested multiple NYC disruption scenarios
What's next
Integrate real-time MTA alerts, historical service disruptions, traffic conditions, and Citi Bike availability to improve temporal analysis and evidence quality.
Built With
- agentic-ai
- elasticsearch
- mistral-ai
- mistral-embed
- python
- semantic-search
- streamlit
Log in or sign up for Devpost to join the conversation.