Human-centered generative AI routing system for public-records automation (prototype for Burnes Center AI for Impact)
This repository contains a demonstration prototype of CivicFlow AI, developed as part of my application for the Generative AI Product Development Fellowship (AI for Impact Program) at the Burnes Center for Social Change, Northeastern University.
The project operationalizes the four-layer framework described in my Candidate Product Specification Memo:
- Trust: Interprets and routes citizen requests accurately.
- Empathy: Generates plain-language, friendly staff responses.
- Equity: Handles multilingual and informal citizen inputs.
- Governance: Includes a human-in-the-loop review and audit system.
City governments often handle thousands of unstructured public records requests via email or web forms. Staff must manually read, classify, and route these requests — leading to delays, backlogs, and inconsistent service.
CivicFlow AI uses a lightweight generative AI model to automatically:
- Interpret request intent.
- Classify it by department, priority, and sensitivity.
- Generate friendly replies.
- Log human corrections for ongoing improvement.
| Layer | Functionality | Implementation |
|---|---|---|
| Trust | Classify and route public requests | Mock AI classifier using realistic pattern rules |
| Empathy | Auto-generate plain-language replies | Rule-based text generator |
| Equity | Handle multilingual (English + Spanish) input | Dataset includes multilingual example |
| Governance | Human approval log | Feedback dataframe and alignment metrics |
| Metric | Result |
|---|---|
| Human–AI Alignment | 100% (simulated) |
| Predicted Departments | 7 |
| Average Response Clarity | Consistent and polite |
| Compliance Incidents | 0 (demo) |
All outputs are synthetic but structured identically to a real civic workflow.
| File | Description |
|---|---|
CivicFlow_AI_Prototype.ipynb |
Main notebook showing the full demo |
data/civicflow_predictions_demo.json |
AI-classified outputs |
data/civicflow_feedback_demo.csv |
Human review log |
docs/Candidate_Product_Specification_Memo_Cherukuru.pdf |
Original memo |
docs/Screenshots/ |
Example outputs and dashboards |
- Integrate with real public records data (via city OnBase or 311 system).
- Fine-tune with staff-reviewed examples to improve model precision.
- Add a Phase 3 citizen-facing portal for self-service search.
- Deploy under Responsible AI and accessibility guidelines.
Swaapnika Chowdary Cherukuru
Master of Science | Northeastern University
cherukuru.sw@northeastern.edu | swaapnic@gmail.com
| https://www.linkedin.com/in/swaapnika-cherukuru-926990228/