Blog Post Award, Global AI Hackathon Series with Qwen Cloud
In rural Thailand, checking a water level shouldn't be a dangerous act — but for some farmers, it is. NaLog Agronomist Agent starts from a sentence that reframes the whole project: no farmer should have to risk their life just to check on water levels. It's a production Qwen memory agent for Thai rice and sugarcane farmers, built on Alibaba Cloud and live today on KhawTECH's NaLog platform — not a demo, a deployed tool serving real communities.
Conventional agritech is priced for industrial farms, not for families who can't afford the devices. NaLog is built for exactly the farmers everyone else overlooks. It gives each one an agronomist with perfect memory that accumulates experience per farmer and per paddy, across sessions and across seasons — remembering what worked in this field last year, what failed, and what the water was doing during flowering.
That memory isn't just storage; it's judgment. The agent detects and supersedes contradictions when new learning conflicts with an existing memory, and it forgets in a timely way — memories decay with age and are physically expired via Tablestore TTL unless they keep proving useful. And it has a safety net with life-or-death importance: a rescue floor ensures that critical memories (reinforcement ≥ 5) always surface in recall results even if they scored below the top-K cutoff, so knowledge like "keep flooded during flowering" is never missed just because the farmer asked an unrelated question.
NaLog is engineered for the hardest deployment conditions imaginable: offline-first, low-bandwidth, rural. It recalls within a tiny context window using 4-tier hybrid recall — vector-first top-K (DashVector plus a qwen3-rerank cross-encoder), BM25-inspired keyword overlap, recency decay, and reinforcement scoring — followed by summarization, so it works even when connectivity is thin.
It also sees the field_: attach a photo and qwen3-vl-plus reads crop condition, water, pests, and weeds directly into the reasoning loop. It acts autonomously but never blindly — sensor alerts trigger unprompted reasoning turns, yet any pump action is only a _proposal that a human must approve before a LoRaWAN downlink is ever sent. And it streams its work: SSE streaming of Qwen's thinking, live tool calls, and the reply, so the farmer literally watches the agent check real sensors before trusting its advice. Each turn includes a memory trace (which memories were recalled, their scores, what was excluded) and a memory diff ("2 new, 1 reinforced, 1 superseded") — transparency as a feature for users who need to trust the machine.
NaLog is a full production deployment on Alibaba Cloud, and the stack reads like a tour of the platform. Qwen models power the reasoning through Model Studio, with qwen3-rerank driving the cross-encoder recall and qwen3-vl-plus reading field photos. DashVector provides the vector layer, Tablestore handles TTL-based forgetting (physically expiring stale memories), Function Compute orchestrates serverless, OSS stores assets, and Qoder supports the development workflow — with DNS and networking completing the deployment. The team published a full proof-of-Alibaba-deployment write-up documenting the live production setup, which is rare for a hackathon project and central to why this one stands out: it's genuinely running in the field.
NaLog was built by Alberto Roura through his agritech startup KhawTECH, with a single, unambiguous mission: to help poor communities in rural Thailand. It's technology aimed not at the easiest market but at the one that needs it most.
Explore the project: https://albertoroura.com/video/nalog-agent-qwen-memory-agent-demo/
Deployment proof: https://github.com/khawtech/nalog-agent/blob/main/docs/proof-of-alibaba-deployment.md
This project was developed as part of the Global AI Hackathon Series with Qwen Cloud.
FORGE: The Hands-Free Voice Copilot for the Technician With Greasy Gloves
Aftershock: The Disaster Simulator That Grades Qwen Agent Societies
148 posts | 29 followers
FollowAlibaba Cloud Project Hub - October 29, 2026
Neel_Shah - February 27, 2025
ApsaraDB - June 4, 2026
VikashThakur - January 3, 2025
Alibaba Cloud Community - March 25, 2022
Alibaba Clouder - October 27, 2020
148 posts | 29 followers
Follow
Token Plan
Build more, spend less. One plan, every modality.
Learn More
QwenWork
QwenWork is dedicated to helping employees strengthen their professional competitiveness in the AI era and to enabling enterprises to improve organizational effectiveness.
Learn More
Qwen
Full-range, open-source, multimodal, and multi-functional
Learn More
Alibaba Cloud Model Studio
A one-stop generative AI platform to build intelligent applications that understand your business, based on Qwen model series such as Qwen-Max and other popular models
Learn MoreMore Posts by Alibaba Cloud Project Hub
