HinaingRadar: Every citizen complaint heard, routed, and resolved on time

Problem Statement

Filipinos complain loudly but are rarely heard systematically. Complaints about uncollected garbage, broken streetlights, flooding, slow permits, and rude staff arrive through the 8888 Citizens’ Complaint Hotline, Facebook comments, text messages, and walk-ins. They are written in Taglish, regional languages, slang, and all caps. Most LGUs triage them by hand, often through one overworked public information officer, so many complaints are lost, duplicated, or forwarded to the wrong department. RA 11032 and the 8888 system expect government offices to act on complaints within set timeframes, yet few LGUs can measure response times. Patterns such as a recurring flood spot or a corrupt window go unnoticed until they make the news.

Proposed Solution

Hinaing AI is a complaint-intake and case management layer for LGUs. Connectors pull in Facebook page comments and messages, an SMS shortcode, web forms, and exported 8888 tickets. A multilingual LLM pipeline then:

• detects the language and normalizes Taglish and Bisaya text
• classifies the complaint by category and responsible office (CENRO, Engineering, BPLO, PNP, barangay)
• extracts the location and geocodes it to a barangay
• scores urgency and sentiment
• clusters duplicates (“47 reports about the same clogged drainage on Rizal St.”)
• writes a one-line summary for the assigned officer

Each case gets an SLA timer with escalation alerts. Citizens receive SMS status updates and an AI-drafted, human-approved reply. A dashboard shows heatmaps, trending issues, sentiment over time, and office-level SLA performance for the mayor.

Target Market

• Primary users: LGU public assistance and complaints desks, PIOs, department heads, and barangay secretaries.
• Buyers (B2G): City/municipal governments, provincial governments, and national agencies with heavy complaint volumes (e.g., utilities regulators, transport agencies). The Office of the President’s 8888 center is a potential integration partner.
• Size: Approximately 1,600+ LGUs and hundreds of national agencies and GOCCs. Large cities receive thousands of social media messages each month.

How AI Is Used

• LLM classification and extraction: Handles code-switched, misspelled, emoji-filled complaints and maps them to the right office and location. Rule-based keyword routing fails on Taglish and sarcasm.
• Embedding-based clustering: Merges duplicate and related complaints into incidents. Without it, staff drown in repeats and miss the scale of an issue.
• Sentiment and urgency scoring: Prioritizes life-safety items such as a downed power line during a typhoon over minor gripes.
• Summarization and reply drafting: Gives busy department heads one-line summaries and polite, Filipino-language reply drafts for human approval.
• Trend detection: Spots spikes (e.g., dengue-related complaints or flood reports) early so the LGU can act proactively.

PLDT / Smart / ePLDT / Rokid Angle

• Smart SMS/A2P messaging: A complaint shortcode and two-way status updates, which is a recurring messaging revenue stream for PLDT Enterprise.
• ePLDT cloud: Secure, locally hosted case data and dashboards for government clients.
• B2G bundles: Sold with LGU connectivity, contact center solutions, and smart city packages.
• Rokid AR (stretch): Field crews such as engineering and garbage teams see the assigned complaint location and photo hands-free while working on site.

Demo Scenario (3 minutes for judges)

1. A screen shows a messy LGU Facebook page: 300 comments in Taglish, Bisaya, and all-caps rants after a typhoon.
2. Click “Ingest.” In seconds Hinaing AI sorts them into categories: Flooding (112), Garbage (64), Power lines (9, flagged URGENT), and Permit delays (21).
3. A cluster card reads: “58 residents report waist-deep flooding at Brgy. San Roque, Purok 3, recurring for 3 days.” It sits on a map pin and is routed to the DRRMO with a 24-hour SLA.
4. A live SMS sent by a judge from their own phone (“wala pa rin kuryente dito sa amin, 2 days na!”) appears, classified and routed in real time.
5. The officer approves an AI-drafted Tagalog reply, and the citizen’s phone receives the status update.
6. The mayor’s dashboard shows SLA compliance by office and a sentiment trend that turns positive after the response.

30-Hour Build Plan

Suggested stack: Python FastAPI, Gemini or Llama 3 / SEA-LION for multilingual classification, sentence-transformers (multilingual) for clustering, PostgreSQL + PostGIS, React with Leaflet and Recharts, a Twilio/Semaphore-style SMS gateway (Smart in production), and the Facebook Graph API (Pages). Public data / APIs: PSGC barangay codes (PSA), OpenStreetMap/Nominatim for geocoding, data.gov.ph, publicly available Filipino sentiment and hate speech datasets on Hugging Face, and team-generated synthetic complaints modeled on public LGU page comments.

Risks & Privacy Notes

• Misrouting or misclassification: Low-confidence cases go to human triage, and officers can reassign cases, which feeds an improvement loop.
• Political misuse (e.g., profiling critics): Strict role-based access, audit logs, and analytics limited to issues rather than individuals in the product design.
• Data Privacy Act 2012: Complainant names and numbers are stored separately and encrypted. The tool processes only what citizens voluntarily submit, follows retention policies, and plans for registration with the NPC as a personal information processor.
• Clause 24 compliance: During the hackathon only synthetic or anonymized complaint data goes into AI tools. No real personal data is scraped or entered.

AI Use Disclosure (draft)

We use an LLM (Gemini or an open model such as SEA-LION/Llama) to classify, extract, summarize, and draft replies, and multilingual sentence embeddings to cluster complaints. AI coding assistants helped write boilerplate backend and dashboard code, and part of our synthetic complaint dataset was AI-generated. Team members reviewed all AI outputs, and every citizen-facing reply requires human approval. No personal or confidential data was entered into AI tools.

Rubric Scoring