🌽 KilimoNova: Agentic USSD for Smallholder Resilience

# Inspiration

In Western Kenya, agriculture is the heartbeat of the economy, yet smallholder farmers face a "Triple Threat": unpredictable climate shifts, devastating pests like the Fall Armyworm, and fragmented supply chains. We were inspired by a simple reality: while AI is booming, it is leaving rural farmers behind. Most AI solutions require $500$ smartphones and 5G—luxuries many in Kakamega cannot afford.

We built KilimoNova to bridge this "Digital Divide." We wanted to move beyond "AI as a chatbot" and create AI as an Agent that lives inside a basic USSD string (*384#). Our goal was to put a PhD-level agronomist and a professional procurement officer into the pocket of every farmer, even on a $20 feature phone.

# What it does

KilimoNova is an autonomous Agentic system that manages the agricultural lifecycle through a simple text interface:

1. Daktari wa Mimea (The Plant Doctor): Powered by Amazon Nova-Micro, this agent processes natural language Swahili reports (e.g., "mahindi yana viwavi"). It diagnoses the threat and provides immediate, localized treatment advice.
2. Afisa wa Dharura (The Emergency Officer): When a critical pest is detected, the agent autonomously triggers a notify_officer tool, sending a high-priority SMS alert to local agronomists with the farmer’s phone number and issue.
3. Wakala wa Pembejeo (Procurement Agent): A transactional closer. Once a pest is identified, the agent facilitates the purchase of the required pesticide, initiating a simulated M-Pesa STK push for a seamless, one-click checkout.

# How we built it

We leveraged the Amazon Bedrock ecosystem to ensure high throughput and low-latency execution:

• The Intelligence: Amazon Nova-Micro serves as the core brain, chosen for its industry-leading inference speed and its ability to handle complex Swahili nuances.
• Orchestration: We used AWS Lambda to manage the state machine between the Africa's Talking USSD gateway and the Bedrock runtime.
• Agentic Tool Use: We implemented function calling within the LLM, allowing the model to decide when to escalate a report to a human officer or when to initiate a payment flow.
• Data Logging: Every interaction is logged as structured JSON in CloudWatch, allowing us to generate real-time "Pest Heatmaps" for NGOs and government monitoring.

# Challenges we ran into

The biggest technical hurdle was "Agentic Verbosity." In a USSD environment, every character counts, and session timeouts are strict (usually $< 10$ seconds). Initially, the LLM would "think out loud" in English, explaining its reasoning before giving the Swahili answer. This caused screen clutter and session timeouts.

We solved this through a pattern we call "Agentic Silence." By implementing a strict Python-based output filter and refined System Prompting, we suppressed the model's internal reasoning blocks. This ensured that only the final, concise Swahili action or advice reached the farmer’s screen.

# Accomplishments that we're proud of

• Zero-UI Accessibility: Building a system that feels "smart" and agentic without requiring a single pixel of a smartphone screen.
• Swahili NLP Precision: Successfully mapping colloquial terms (like "viwavi") to scientific pest categories and urgent tool-triggers.
• The Advice-to-Action Loop: Successfully handing off a "Diagnosis" to an "Officer Alert" and a "Payment Request" in under $60$ seconds.

# What we learned

We learned that in the African context, Deployment is as important as Intelligence. An AI model that requires a high-end UI is useless to a farmer in a remote village. By focusing on "Agentic" action over "Generative" fluff, we created a tool with immediate economic utility.

Mathematically, we modeled our impact using the Response Velocity ($V$):

$$V = \frac{R}{L}$$

Where:

• $R$ is the Resource Mobilization (sending the officer/ordering medicine).
• $L$ is the Latency of detection.

By using Nova-Micro to minimize $L$, we maximize $V$, effectively stopping pest outbreaks before they become regional disasters.

# What's next for KilimoNova

We plan to integrate Amazon Nova Reel to generate 6-second instructional videos in local dialects, sent via SMS links, showing farmers how to apply the pesticides they just bought. We also aim to partner with micro-lending institutions to allow the agent to apply for "input loans" on behalf of the farmer directly within the USSD session.

Built With

• amazon-nova
• aws-lambda
• python
• ussd