Serbisyo Copilot: Know exactly what to bring, before you line up at city hall

Problem Statement

Many Filipinos lose a full day’s wage going back to city hall two or three times because a requirement was missing, outdated, or unreadable. Requirements differ from LGU to LGU. They are often posted on faded tarpaulins or scattered Facebook posts, and they are usually in English legalese. Republic Act 11032 (Ease of Doing Business and Efficient Government Service Delivery Act) requires processing within 3, 7, or 20 working days and a published Citizen’s Charter. In practice, frontline staff still spend much of their time answering the same questions and rejecting incomplete submissions. The Philippines has approximately 1,600+ cities and municipalities and 42,000+ barangays, and most have no digital pre-screening at all, which keeps queues long and fixers in business.

Proposed Solution

Serbisyo Copilot is a web/Messenger chatbot and kiosk app. Citizens describe what they need in their own words (“kukuha ko og business permit para sa sari-sari store”). The copilot answers using a retrieval-augmented knowledge base built from each LGU’s official Citizen’s Charter and returns a personalized checklist with fees, office location, and processing time. Citizens then photograph their documents. A vision model classifies each one (e.g., DTI certificate, barangay clearance, valid ID, lease contract), checks expiry dates, signatures, and legibility, and flags anything missing. Before upload, an on-device model blurs ID numbers, faces, and addresses that the check doesn’t need. The citizen gets a “Ready to Submit” QR code that the frontline officer scans, so a 10-minute verification at the counter becomes a 1-minute confirmation.

Target Market

• Primary users: Citizens and micro-entrepreneurs dealing with LGU services: cedula (community tax certificate), barangay clearance, business permits and renewals (peak season is January), building and occupancy permits, and civil registry requests.
• Buyers (B2G): City/municipal governments, BPLOs, and barangays. These offices are under ARTA compliance pressure and DILG Seal of Good Local Governance scoring, and potential champions include ARTA and DICT (eGov PH).
• Size: Approximately 1,600+ LGUs, with tens of millions of transactions per year. Even mid-sized cities process tens of thousands of business permit renewals every January.

How AI Is Used

• LLM with RAG over Citizen’s Charters: Understands Taglish, Bisaya, and Ilocano questions and maps vague intents to the exact service and requirement list. Keyword FAQ bots break on code-switching and on the wide variety of ways citizens phrase a request.
• Vision document classification and checking (OCR plus a multimodal model): Identifies document types, reads expiry dates, and detects blur, cropping, or missing pages. Without it, completeness checks stay manual, which is the main bottleneck.
• On-device PII redaction (lightweight detector, e.g., a small YOLO/OCR model in the browser): Masks sensitive fields before anything leaves the phone. Without it, LGUs could not adopt the tool under the Data Privacy Act.
• Analytics summarization: Shows LGUs which requirements cause the most rejections, so they can simplify their charters.

PLDT / Smart / ePLDT / Rokid Angle

• ePLDT cloud and data centers: Sovereign, local hosting of LGU knowledge bases and models, which matters for government data residency.
• Smart SMS/Messenger connectivity: SMS fallback for checklists and “ready for pickup” notifications for citizens without data plans.
• PLDT Enterprise B2G sales: Can be bundled with LGU connectivity, queue kiosks, and digital transformation packages.
• Rokid AR glasses (stretch): Frontline officers glance at a scanned QR and instantly see a verified checklist overlay, keeping their hands free for stamping.

Demo Scenario (3 minutes for judges)

1. Aling Nena, who runs a carinderia in Cebu, has already been sent home twice from city hall. She types in Bisaya: “Unsay kinahanglan para mag-renew sa akong business permit?”
2. The copilot replies in Bisaya with a 6-item checklist, fees, and the BPLO window number, citing the LGU charter.
3. She photographs her documents. The app flags that her barangay clearance has expired and her lease contract is missing page 2, and her ID number is auto-blurred on screen.
4. After she re-uploads, everything turns green and a “Ready to Submit” QR appears.
5. Switch to the officer view: the QR is scanned and verified in seconds.
6. Close with an LGU dashboard showing that “expired barangay clearance” causes 31% of rejections in the simulated data.

30-Hour Build Plan

Suggested stack: Next.js PWA, Gemini or an open multimodal model (e.g., Qwen-VL) for document understanding, Tesseract/PaddleOCR, Transformers.js or ONNX Runtime Web for on-device redaction, a vector DB (pgvector), Supabase, and Messenger Platform API. Public data / APIs: LGU Citizen’s Charters (required by ARTA and published on LGU websites), ARTA resources, data.gov.ph, PSA OpenSTAT (LGU population for sizing), Hugging Face multilingual models (Tagalog/Cebuano), and synthetic sample documents generated by the team.

Risks & Privacy Notes

• Hallucinated requirements: Answers are grounded only in retrieved charter text with citations, and the bot says “please confirm with BPLO” when it isn’t confident.
• Fake or forged documents: The tool checks completeness, not authenticity. Final verification stays with LGU officers.
• Data Privacy Act 2012: On-device redaction, data minimization, consent screens, auto-deletion after the transaction, and a Privacy Impact Assessment before any pilot.
• Clause 24 compliance: No real citizen IDs or personal data are entered into AI tools during the hackathon. Only synthetic or mock documents are used.

AI Use Disclosure (draft)

We use Gemini (or an open-source equivalent) for multilingual question answering and document understanding, PaddleOCR for text extraction, and a small on-device model for PII masking. AI coding assistants helped scaffold parts of the frontend, and synthetic test documents were generated with AI. Every AI-generated output, including code, checklists, and sample documents, was reviewed by team members. No personal or confidential data was entered into any AI tool.

Rubric Scoring