In the rapidly evolving landscape of Artificial Intelligence, we are moving away from simple “chatbots” toward autonomous agents—systems that can perceive the world, reason over time, and recommend complex actions. But how do these agents perform in the high-stakes, “perma-crisis” environment of a megacity in the Global South?
I am honored to share that our 100-day longitudinal study on this topic was recently accepted and presented at the 2nd International Conference on Human-AI Interaction and Experience Design (HAXD 2026), an IEEE technically co-sponsored event in Valencia, Spain.
The Paper: Designing Autonomous Agents for Urban Crisis Management: A Longitudinal Evaluation of LLM Reasoning in the Global South
Authors: Taniv Ashraf, Suniv Ashraf, Saidul Ashraf
Venue: IEEE HAXD 2026
The Urban Challenge: Signal vs. Noise
Cities like Dhaka, Bangladesh, face a unique set of challenges: hazardous air quality, chronic traffic congestion, and sudden safety shocks like earthquakes or public health outbreaks. For human decision-makers, the problem isn’t a lack of information—it’s information overload. Thousands of news signals are generated daily in Bengali, making it nearly impossible to prioritize the most salient threats in real-time.
Introducing Project Jagoron (Awakening)
To address this, I architected Project Jagoron. This is not just a prompt; it is a fully automated agentic pipeline designed to act as a 24/7 strategic sentry. The system follows a rigorous daily cycle:
- Perception: Ingests 50+ Bengali news RSS feeds using Python.
- Cognition: Processes raw text through Google Gemini 1.5 Pro to filter noise and identify life-safety priorities.
- Action: Generates a structured “Strategic Briefing” in Bangla and stores it in a Supabase database.
- Transparency: Pushes results to a live web application hosted on Vercel.
The Discovery: Dynamic Persona Shifting
Over our 100-day evaluation (monitoring ~50,000 words of AI advice), we made a fascinating discovery that the HAXD committee highlighted: Emergent Persona Adaptability.
We found that the agent autonomously shifted its communicative tone based on the crisis typology:
- The Manager Persona: During routine infrastructure failures (e.g., Metro Rail delays), the AI adopted a technocratic, data-driven tone.
- The Humanitarian Persona: During life-threatening events (e.g., seismic tremors), the AI autonomously shifted to an empathetic, people-centered posture, suggesting mental health support and visible relief efforts.
This suggests that modern LLMs, when placed in a longitudinal agentic workflow, can become “situationally aware” partners that align with human safety needs.
Quantitative Results
Using Google Colab to analyze the output, our data showed that the AI consistently prioritized Public Health (61%) and Political Instability (49%) as the most salient issues for governance. This proves that an autonomous agent can successfully filter a chaotic news cycle to focus on the topics that directly impact human survival.
Next Steps: Human-in-the-Loop
As noted during our Q&A session at the conference, the next phase of this research is “Human-in-the-Loop” validation. We intend to present these AI-generated briefings to actual city administrators and urban planners in Dhaka to measure the practical utility of this “Robot Strategist” in real-world governance.
Explore the Project
We believe in open science and transparency. You can explore the entire ecosystem below:
🚀 Live App: Project Jagoron Briefing Archive
💻 Source Code: GitHub Repository
🤝 Connect: Taniv Ashraf on LinkedIn
Special thanks to the HAXD 2026 organizing committee and Dr. Addi Ait-Mlouk for the valuable feedback during the peer-review process.