Blog Post Award — Global AI Hackathon Series with Qwen Cloud

Project Overview

In rural Thailand, checking water levels should never be dangerous—but for some farmers, it actually is. The NaLog Agricultural Engineer Agent redefines the perspective of the entire project through a single statement: “No one should have to risk their life to check the water level.” This project is a Qwen memory agent deployed in production environments for rice and sugarcane farmers in Thailand. Built on Alibaba Cloud, it currently operates in real time on KhawTECH’s NaLog platform. This is not a demo but a practical tool deployed to genuinely support local communities.

Project Value

Conventional agritech solutions are priced for large-scale farms and are out of reach for small family-run farmers who cannot afford to purchase equipment. NaLog, on the other hand, is designed for precisely those farmers whom no one else is paying attention to. It provides each farmer with a “virtual agricultural engineer” with perfect memory, accumulating experience across sessions and seasons for each farmer and each paddy field. For example, it accurately remembers all kinds of information, such as “the management practices that worked in this plot last year,” “cases that failed,” and “how water levels changed during the flowering period.”

This memory is not merely data storage; it also includes decision-making capabilities. When a new learning result conflicts with existing memory, the agent automatically detects the contradiction and supersedes the old memory. Memories are also “forgotten” at the appropriate time. In other words, memories decay over time, physically expire, and are deleted through Tablestore’s TTL (Time-To-Live) feature—except when a memory has consistently proven useful and is retained as an exception. Furthermore, as a critical life-and-death safety mechanism, there is a “rescue floor” (minimum rescue threshold). This guarantees that extremely important memories with a reinforcement score of 5 or higher—such as “maintain flooding during the flowering period”—are always included in search results, even if they fall outside the top-K candidates. Therefore, even if a farmer asks an unrelated question, life-saving knowledge will never be overlooked.

Solution

NaLog is designed for what may be the harshest deployment conditions imaginable—offline-first operation, low bandwidth, and remote rural areas.To enable fast information retrieval even within an extremely narrow context window, NaLog employs a four-stage hybrid retrieval (hybrid recall):

1. Top-K retrieval using vector search (DashVector + qwen3-rerank cross-encoder),
2. keyword-overlap-based retrieval applying BM25,
3. recency decay based on time series,
4. reinforcement scoring.
   It then summarizes the search results, enabling reliable operation even in unstable communication environments.

Furthermore, NaLog can “see the field.” When a photo is attached, qwen3-vl-plus directly reads the crop growth status, water level, pests, and weeds, and immediately incorporates that information into the reasoning process. The agent acts autonomously, but never acts thoughtlessly. Alerts from sensors automatically initiate reasoning without instructions from the user, but execution actions such as pump operation remain “proposals” and always require human approval before actual commands are sent via LoRaWAN. The processing process is also visualized in real time. Through Server-Sent Events (SSE), Qwen’s thought process, the execution status of tool calls, and the final answer are streamed in sequence, allowing farmers to visibly confirm that the agent is actually checking the field sensors. Each reasoning step is accompanied by a “memory trace” (which memories were retrieved, their respective scores, and the memories that were excluded) and a “memory diff” (e.g., “2 new memories, 1 reinforced, 1 overwritten”), making transparency a feature in itself for users who require reliability.

Main features

NaLog is deployed as a fully operational environment on Alibaba Cloud, and the technology stack in use is like a “tour” that showcases the comprehensive utilization of the platform. Through Model Studio, Qwen models power inference; qwen3-rerank handles cross-encoder retrieval; and qwen3-vl-plus handles recognition of field photos. DashVector provides the vector search layer, Tablestore implements TTL-based “forgetting,” Function Compute orchestrates serverless processing, OSS stores asset files, and Qoder supports development workflows. DNS and networking configuration complete the deployment. The team has published an “Alibaba Cloud Implementation Certificate (Proof of Alibaba Deployment)” documenting the actual operating environment in detail. This is an extremely rare undertaking for a hackathon project and the decisive reason this project stands apart from the others—namely, NaLog is truly operating in the field.

Alibaba Cloud products used

• Qwen Cloud
• Model Studio
• Qwen models
• OSS
• Qoder
• Others: DashVector, TableStore, DNS, Function Compute

About the Development Team

NaLog was developed by the agritech startup “KhawTECH,” led by Alberto Roura. Its mission is clear and unique—supporting low-income communities in Thailand’s provincial areas. This technology is aimed not at the easiest market, but at those who need it most.

Watch the project introduction video: https://albertoroura.com/video/nalog-agent-qwen-memory-agent-demo/
Evidence of a production deployment: https://github.com/khawtech/nalog-agent/blob/main/docs/proof-of-alibaba-deployment.md
This project was developed as part of a global AI hackathon series with Qwen Cloud.

This article is a translation of an Alibaba Cloud Blog article.