The Register wrote about HarvestBench. In a farming simulation, most LLM agents will drive a tractor through an animal rather than spend extra fuel going around it, and the spread between models is enormous. GPT-5.6 Terra killed animals in 0.4% of contacts. GPT-4o mini did in 98.8%. Three things we found that matter beyond farms: Telling the model it is being scored on "morality" changed behaviour a lot. Removing that one line changed kill rates from under 1% to over 80%. Prompted 'values' are too fragile to rely on. Nearly every model spared farmed animals more often than wild ones. The probably interpretation is that models value animals by their economic value to the farmer/ human, not by the animal's own interests. Models that say 'a pig matters' when asked directly will still run one over when it costs fuel. What a model says and what it does in an agentic setting have different answers. Paper: arxiv.org/abs/2609.04444 Live board: https://lnkd.in/g5Br5qT3 Code: https://lnkd.in/ge3pCkAF Article: https://lnkd.in/gzi7fXna With Miles Tidmarsh, Matthias Endres, Anshuman Singh and Jeremiah Miller at Compassion Aligned Machine Learning, with the University of Warwick. Built on Inspect from the UK AI Security Institute (AISI).

🚀 Moving Beyond Jupyter Notebooks: Building an Event-Driven AI Proxy Agent for B2B Animal Rescue Networks 🐾 Many data scientists get stuck optimizing model metrics on static CSV files inside isolated environments. But in production engineering, value is unlocked when you transform those predictive models into autonomous, closed-loop systems. Leveraging historical datasets and modeling insights from the Kaggle PetFinder Pawpularity Challenge, I built and deployed a production-grade Event-Driven AI Proxy Agent designed to intercept and optimize pet profile listings mid-flight before they are broadcast across global syndication feeds (like RescueGroups.org or Shelterluv). 🏗️ How the System Design Works: Instead of forcing shelter workers to manually edit photos, my system acts as an automated middleware layer inside a REST API: 1.Payload Interception: A web gateway built with FastAPI catches raw incoming JSON uploads, parsing binary base64 imagery metrics on the fly. 2.The Actor (Computer Vision): The agent invokes an embedded YOLOv8 object detection backbone to locate the precise spatial coordinates of the pet, automatically executing a smart bounding-box macro-crop. 3.The Critic (Deep Learning): The cropped image is fed through an EfficientNet-B0 classifier (trained via the timm PyTorch library), utilizing a normalized binary cross-entropy sigmoid head to rank the asset's structural appeal. 4.Closed-Loop Optimization: The agent recursively executes micro-exposure/contrast adjustments via OpenCV, checking its work against the Critic until the image passes quality parameters before letting the finalized payload proceed to the global broadcast stream. 🛠️ Key Takeaways & Architecture Focus: .Production-First Mentality: Transitioned a traditional ML regression challenge into a modular, containerized web service. .Network Serialization: Handled dynamic matrix array transformations and base64 transmission pipeline optimizations. .Multi-Model Orchestration: Successfully synchronized a deep learning critic framework with real-time geometric computer vision filters. Building this end-to-end framework highlights exactly how we can bridge advanced predictive algorithms with scalable software system design patterns. 📁 Check out the complete codebase, installation instructions, and system flow diagrams here on my GitHub repository: 👉 [ https://lnkd.in/erXDCQjm ] RescueGroups.org #MachineLearning #AI #SystemDesign #Python #ComputerVision #PyTorch #FastAPI #MLESprit #DataScience #SoftwareEngineering

Excited to share LivestockGuardian Vision — a local-first, production-style livestock monitoring prototype built with Python, OpenCV, Streamlit, and SQLite. It runs entirely on your machine with no API key required and includes: - An original synthetic 30–60 second simulation video generated locally, with scripted scenarios such as normal movement, occlusion, entry/exit, zone violation, herd separation, inactivity, gate crossing, and feeding behavior - Multi-object tracking with persistent IDs, centroids, bounding boxes, confidence, track age, hits, misses, and trajectory history - Editable ROIs, pens, feeding zones, water zones, and virtual gate lines - A rule-based event engine with debounce/cooldown, evidence frames, and human-readable explanations - Smoothed velocity, direction estimation, and short-horizon trajectory prediction with uncertainty overlays - SQLite persistence for sessions, tracks, track points, zones, events, and metrics - CSV/JSON export and a local SQL-backed natural-language query layer - Simulation evaluation against ground truth: counting error, ID switches, fragmentation, position error, event precision/recall, and FPS The key design principle: no fake domain claims. The simulation uses deterministic ground-truth detection, while uploaded/webcam video uses a clearly labeled generic motion detector. Built as a portfolio-grade prototype for AI-based livestock monitoring, edge analytics, and simulation-first evaluation. #ComputerVision #MultiObjectTracking #LivestockMonitoring #AI #OpenCV #Streamlit #Python #MachineLearning #AgriTech #EdgeAI

🌾 From a Problem Statement to a Working Prototype — AgriAI 24 hours. One problem statement. A lot of debugging. And a working prototype. 🚀 We recently participated in the PRARAMBHA 2.0 – 24 Hour Online Hackathon, organized by D.K.T.E. Society's Institute of Engineering and Technology, Ichalkaranji. Our challenge was to build a solution around AI-Powered Digital Farm History. That’s where we built AgriAI — a digital farm record system designed to help farmers keep track of their farming journey and ask questions about their own records. With AgriAI, farmers can maintain: 🌱 Farm & field details 📅 Seasons & crops 🚜 Farming activities 🧪 Fertilizers and other inputs 💰 Expenses 🌾 Harvest records 🤖 AI-powered questions about their farm history 👨💻 My Role — Backend & AI Integration I worked primarily on the backend development and AI integration of AgriAI, including: 🔹 Designing and implementing the FastAPI backend 🔹 Building the PostgreSQL database structure for farm records 🔹 Developing APIs for farms, fields, seasons, crops, activities, inputs, expenses and harvests 🔹 Implementing authentication and JWT-based authorization 🔹 Building the AI query pipeline to retrieve relevant farm records before generating an answer 🔹 Implementing evidence-backed AI responses connected to actual farm data 🔹 Working on Android ↔ FastAPI integration and testing the system on a physical device The interesting part was making the AI work with actual farm records rather than simply generating generic answers. For example: “How much Urea did I use?” AgriAI retrieves the relevant farm record and provides an answer along with the supporting evidence. Building it in 24 hours wasn't easy. We faced API integration issues, database challenges, Android connectivity problems, AI API limitations, and plenty of debugging 😅 But that's what made the experience valuable. We didn't just build an app — we experienced what it takes to turn an idea into a working product under a deadline. A big thanks to my team members for the collaboration and to D.K.T.E. Society's Institute of Engineering and Technology, Ichalkaranji for organizing the hackathon and giving us the opportunity to build, experiment, and learn. 🙌 🚀 One problem. One team. 24 hours. One working prototype. #AgriAI #PRARAMBHA2 #PRARAMBHA20 #Hackathon #DKTE #Agritech #ArtificialIntelligence #AI #AndroidDevelopment #FastAPI #Python #PostgreSQL #GenerativeAI #BackendDevelopment #SoftwareDevelopment #HackathonExperience #Innovation #StudentDeveloper

I promised myself that I was going to get serious about data analysis, so here's another proof. Alongside my group members, I worked on a linear programming case study tackling a real agricultural decision: a farmer has 20 hectares of land, 860 working men, and ₦3.5 million to work with. He wants to plant cassava, maize, and vegetables but how should he divide his land and resources across the three to maximize profit? This is where data stops being abstract and starts being useful. Using linear programming, we modeled the constraints (land, labor, budget) against the goal (profit maximization) to find the optimal allocation , how many hectares should go to each crop, and how resources should be distributed to get the best possible return. This is exactly the kind of thinking I want to bring more of into agriculture: decisions grounded in data, not guesswork. Farming already carries enough uncertainty from weather and markets, the decisions we can control, like resource allocation, deserve the same rigor as any business problem. The last two pages of the attached document show us working through the analysis in Excel. Shoutout to my amazing group members: Josephine Umoru Victor Olatunbosun and others this was a true team effort. I'm Benedicta Ayeni an animal science student and business data analyst passionate about bringing data-driven thinking to agriculture. #Agriculture #DataAnalysis #LinearProgramming #Agribusiness #LivestockProduction #OAU

John Deere is launching an AI assistant to help farmers make sense of their field data. The tool, built into Deere's Operations Center mobile app, draws on field data to help guide decisions on fertilizer use and other choices. "We think there's hundreds of decisions that economically matter for a farmer throughout a season," Deanna Kovar, worldwide president of agriculture and turf at Deere, told Bloomberg. "JD is now the researcher for the farmer — the data analyst — to go seek out those answers and do it in a much more efficient way."

Dr. Karen Hildebrand, PhD of Amazon Web Services (AWS) discusses data centers with the guys. The need for them but also the need for people to voice their opinions. 🎧 Listen to the full episode: https://bit.ly/4xiFgXD #Farm4Profit #Farm4Fun #Agriculture #Farming #AgTech #ArtificialIntelligence #AIinAgriculture #PrecisionAgriculture #SmartFarming #FarmTechnology #FutureOfFarming #Agribusiness #AWS #Harvest2026 #Harvest26 #TechHubLive #datacenter

This helped us win our first hackathon 🏆 What started as an overnight hackathon problem statement turned into PashuPehchan, an AI-assisted livestock platform built around a simple idea: AI suggests. Humans verify. We built a system to help with cattle and buffalo breed identification, while connecting that workflow to actual farmer needs. 🐄 What we built AI-based cattle & buffalo breed identification Top-3 breed predictions with confidence scores Human verification of AI predictions Digital livestock records Farmer portal Separate Middleman marketplace portal Livestock listings, search, comparison and enquiries Nearby veterinary service discovery Breed library and verification history ⚙️ Tech stack Frontend: React + Vite Backend: Python + FastAPI Database: SQLite ML: PyTorch + EfficientNet-B0 Deployment: Vercel For the ML pipeline, we used Indian bovine breed datasets from Kaggle and Hugging Face. We used the structured Hugging Face dataset for training/validation/testing and fine-tuned a pretrained EfficientNet-B0 model for breed classification. The model then feeds into our FastAPI backend, where an uploaded image is processed and the system returns the top-3 predicted breeds + confidence, which the user can verify before saving the animal record. The biggest challenge wasn't just getting a model to classify images — it was turning that model into an actual usable workflow for farmers. 🌐 PashuPehchan: https://lnkd.in/dKGx4K5k #Hackathon #OriginHackathon #PashuPehchan #AI #MachineLearning #ComputerVision #PyTorch #FastAPI #React #AgriTech #FarmTech #LivestockTech #ArtificialIntelligence #BuildInPublic

Funny, a lot of the accounts here and on Twitter haven't received the memo Postgres and Graphs are popular for engagement farming. Most articles on the topic tend to be shallow and regurgitate famous platitudes without much thought. Just use Postgres. You don't need Graph Databases.

Pets vs Cattle! It's the 2010's again but now with code instead of infrastructure! The "Pets vs Cattle" mental model switch was a very exciting time to me. The most exciting part of my career so far were spent around CI/CD automation and observability. I've been lucky to attend the first Monitorama conference in 2013 in Boston and the penultimate one in Portland in 2022. With the need of scaling out services, the industry had to reinvent everything. At first, configuration management tools like Puppet, Chef, or CFEngine hold the most important part of the conversation, and then came VM automation, containers, orchestrators, and observability. We're now in 2026 and software writes a huge chunk of the code. The majority of code will be cattle. A few very specific one will be pets. The code hosting infrastructure isn't ready for this switch either. It has been built for the "code as pets" era but we're entering the "code as cattle" era. Take a look at the slide below. It's from 2012. The first thing about pets is giving them a cute name. Now create a new repository on GitHub. What's the first thing they'll ask you? A name! How cute!