The problem

Bhutan has made meaningful progress in digital governance, yet citizens, especially in rural areas, still face fragmented, manual, and inaccessible public service workflows. A single service request can require navigating multiple government offices, filling out forms in multiple systems, and waiting weeks for manual processing.

Language is a compounding barrier. Existing digital platforms primarily serve English-literate users, while a significant share of the population communicates in Dzongkha or other local languages. Low digital literacy further widens the gap between available services and those who can actually access them.

On the government side, staff must manually process large volumes of requests across disconnected systems, creating delays and inefficiencies that scale poorly as service demand grows.

This raises critical questions:

• Can conversational AI reliably understand citizen requests in both English and Dzongkha?
• Can agentic workflows automate cross-agency coordination while remaining transparent and explainable?
• Can voice-first interfaces serve citizens with limited digital literacy in real-world conditions?

Without validating these questions, AI cannot responsibly scale in the Bhutanese public sector.

The project goals

This challenge builds CivicAI Bhutan, an experimental MVP platform that explores the feasibility of voice-enabled, multilingual, agent-based public service delivery for Bhutan’s citizens.

The project focuses on:

• Experimenting with multilingual NLP for Dzongkha and English intent detection.
• Building agentic workflows that route requests across simulated government agencies.
• Developing voice interaction pipelines (ASR + TTS) and evaluating Dzongkha coverage.
• Implementing RAG-grounded conversational AI using government policy documents.
• Integrating explainability mechanisms for every automated decision step.
• Combining validated components into a functional MVP prototype.
• Technical documentation.

The system must demonstrate the ability to:

• Accept citizen requests via voice or chat in English or Dzongkha.
• Classify intent and route requests to the correct simulated agency workflow.
• Automate multi-step processes with transparent, traceable decision logs.
• Gracefully degrade from voice to text when voice input is unavailable.
• Identify failure cases, coverage gaps, and edge scenarios honestly.

Impact of the Problem

AgriVision Global can directly impact how agriculture operates on the ground, especially in regions where access to reliable diagnostics is limited.

Citizens

• Faster service access through natural voice or chat, no digital literacy required. Requests understood in Dzongkha.

Government staff

• Reduced manual request processing through automated routing. More capacity for complex cases requiring human judgment.

Rural communities

• Voice-first design and Dzongkha support directly address the accessibility gap for citizens in remote areas.

Civic AI ecosystem

• First validated dataset and architecture for Dzongkha NLP in a public service context, reusable across Bhutan’s future AI initiatives.

Timeline

1

Sprint 1: Foundation, work streams & multilingual data baseline (Weeks 1-2). 

Review Local Chapter outputs, define 5 parallel work streams, set up shared repositories, and evaluation rubrics. Source and synthesize Dzongkha + English datasets. Establish a multilingual NLP baseline using XLM-R and NLLB-200.

2

Sprint 2: NLP classification & RAG-grounded conversational AI (Weeks 3-4).

Experiment with intent detection across Dzongkha and English inputs. Build RAG pipeline using government policy documents. Test chain-of-thought prompting for transparent decision traces. Benchmark classification approaches.

3

Sprint 3: Agentic workflows & voice interaction experiments (Weeks 5-6).

Build multi-agent routing workflows (LangGraph / ReAct) across simulated agency endpoints. In parallel: test Whisper ASR + Silero VAD pipeline, evaluate Dzongkha coverage, and prototype TTS. Measure end-to-end latency and usability.

4

Sprint 4: MVP integration, evaluation & documentation (Weeks 7-8)

Combine the strongest components into a single citizen-journey prototype. Conduct end-to-end testing. Produce the final technical report and Dzongkha data gap analysis.

More details will be shared with the designated team.