[Public Service Innovation] Saving 500 Hours Monthly with "Offline AI" & Enhancing Citizen Service via Data Insights

As the Government actively promotes the strategic vision of a "Smart Government" and "AI Efficiency Enhancement," departments are vigorously seeking innovative technologies to optimize public service workflows. We understand that true digital transformation is not merely about introducing new tech, but seamlessly integrating it into frontline operations to achieve "efficiency and speed" without sacrificing service quality.

We recently developed a cutting-edge AI solution for a leading local airline. This technology not only solved operational efficiency issues in a high-pressure environment but also serves as an excellent reference case for government frontline services (such as service counters, hotlines, and triage stations).

1. Primary Benefit: Unlocking Workforce Capacity & Efficiency

In this airline use case, we replaced traditional manual data entry with AI voice technology. By removing the cumbersome process of navigating through sub-menus, we successfully saved approximately 20 seconds per interaction.

While 20 seconds may seem minor, the cumulative effect is substantial:

• Hours Released: The project is projected to save 500 worker hours per month for the operation, significantly reducing manpower costs.
• Process Optimization: It liberates frontline staff from the role of "data entry clerks," allowing them to handle higher-value service requests.

2. User Experience: Breaking the "Screen Wall"

While pursuing efficiency, we are also dedicated to solving a common side effect of digitization: "The Screen Wall". Traditionally, frontline staff are forced to break eye contact with citizens to focus on tablet operations, making the service feel impersonal.

Our solution creates a "Heads-Up" service experience. Through voice commands, the system automatically populates data in the background, allowing staff to maintain eye contact throughout the interaction. This "Invisible UI" ensures citizens feel "heard" rather than "processed", perfectly aligning with a people-centric service philosophy.

3. Data Insights: From Interaction to Strategic Optimization

Beyond solving immediate service needs, this system possesses powerful potential for backend analysis. By automatically analyzing accumulated conversation content, we can transform frontline interactions into concrete management data to develop further use cases:

• Hotspot Analysis & Knowledge Base Optimization: The system can identify and catalogue the most frequently asked questions, assisting departments in updating FAQs or training materials in real-time, and can even suggest standardized answers to staff to ensure consistency.
• Sentiment Analysis: By analyzing tone and word choice, the system can objectively evaluate citizen satisfaction trends, helping management gauge public sentiment.
• Red Flag Detection: The system can be configured to detect specific sensitive keywords or abnormal interaction patterns, providing early risk warnings for potential complaints or emergencies, enabling proactive management intervention.

4. Technical Challenges: Overcoming the "Hong Kong Context"

To implement this technology in a real-world Hong Kong environment, we had to overcome two major pain points that standard off-the-shelf AI models could not solve:

• Challenge 1: The "Hong Kong Slang" Factor
  Standard Speech-to-Text (STT) engines struggle to process Hong Kong's unique mix of Cantonese and English (Code-mixing) and local terminology (e.g., specific requests mixed with slang). Generic models often hallucinate or fail to transcribe accurately.
• Challenge 2: Data Privacy & Network Latency
  For services involving sensitive data, uploading voice data to the cloud poses privacy risks. Furthermore, network latency can cause service interruptions, which is unacceptable in time-critical frontline environments.

5. The Solution: Localization & On-Device AI

To address these challenges, we implemented targeted technical breakthroughs:

• Localized Fine-Tuning: Instead of using a generic model, we performed specific fine-tuning on the STT engine to accurately recognize complex instructions containing local slang and mixed languages.
• On-Device AI: We successfully optimized Large Language Models (LLMs) to run offline on mobile devices. This means all data is processed locally on the device without needing the internet, thoroughly resolving data leakage concerns while achieving zero-latency response.

We believe this solution, combining "Data Insights," "High Privacy," and "Localized Recognition," can effectively support your department in realizing the smart upgrade of public services.