Auctus AI · Agricultural intelligence
A working AI lab inside an eighteen-year agribusiness practice. Custom model training, AI-assisted compliance, and agronomic agents — built by people who have walked the rows and read the regulations. African agribusiness AI work that takes the agronomy as seriously as the model.
Four streams
Auctus AI delivers across four technical streams. Most engagements draw on two or three of them in combination, scoped to the brief.
Crop disease detection, weed and pest identification, plant-health monitoring, and stand-count estimation from drone, satellite, and phone-camera imagery. Trained on labelled field data the firm collects in-country.
Frameworks: PyTorch, TensorFlow, Hugging Face Vision. Open-source backbone models (YOLO, ResNet, ViT) fine-tuned on agricultural datasets.
Yield prediction, input-cost optimisation, price and commodity forecasting, post-harvest loss modelling, irrigation scheduling. Combines on-farm sensor data, three to five seasons of historical records, climate inputs, and market data.
Methods: Gradient-boosted trees (XGBoost, LightGBM), time-series models (Prophet, classical ARIMA, deep learning where data depth supports it), and explainable ML (SHAP) so the operations team can read why the model said what it said.
AI-assisted compliance documentation, regulatory surveillance, audit-prep automation, and bilingual extension content. Built on retrieval-augmented generation (RAG) over the GlobalG.A.P. v6 standard, EU MRL database, EU–SADC EPA texts, and SPS protocols.
Stack: OpenAI, Anthropic Claude, and open-source models via Hugging Face (Llama, Mistral). LangChain / LlamaIndex for retrieval pipelines. Fine-tuned on agricultural domain language where the use case warrants it.
Multi-step decision-support agents for agronomic Q&A, spray-window recommendation, compliance walkthroughs, and market-intelligence briefings. Agents reason over the firm's knowledge base and the client's operational data, with human-in-the-loop validation built in.
Architecture: Tool-using agents with structured outputs, supervised by a senior agronomist on every recommendation that touches a regulatory or commercial decision.
Use cases · in development or in production
Phone-camera and drone-acquired imagery trained to flag early-stage citrus black spot (CBS) and phytophthora root rot. Reduces scout time and surfaces problem blocks before yield impact is visible at harvest.
Block-level yield prediction combining three to five seasons of records with weather inputs and orchard-level sensor data. Output feeds the Power BI dashboards built under Precision Agriculture engagements.
RAG-based assistant trained on the v6 standard plus the cooperative's quality management system. Drafts spray records, internal-inspection reports, and worker-induction documentation in literacy-appropriate language for auditor review.
Continuous monitoring of EU pesticide residue database changes for active ingredients in the client's spray programme. Alerts the operations team to scheduled MRL reductions twelve months ahead of effective date.
Question-answering agent for smallholder cooperative members on agronomic best practice, in English and selected SADC languages. Trained on regional extension content plus the firm's accumulated field knowledge. Operates by phone (USSD/WhatsApp) where smartphone penetration is partial.
Agronomic decision support for commercial hemp and medicinal cannabis operations. Combines varietal data from in-house breeding research with on-farm sensor inputs to recommend irrigation, nutrition, and IPM adjustments at growth-stage granularity.
Stack
The lab uses open-source frameworks where reasonable, with commercial models where they outperform. Every model is documented for reproducibility; every recommendation is traceable to the data that produced it.
Operating principles
A model that doesn't match the agronomic question is wasted compute. Every Auctus AI engagement starts with the agronomic problem, not the technical method.
No AI output that affects a compliance audit, a spray decision, or a buyer commitment goes to the client without a senior agronomist's review. The model accelerates the human; it does not replace them.
A 1% improvement in accuracy that the operations team can't read is worth less than the previous model the team trusts. SHAP, attention maps, and traceable inputs by default.
Llama and Mistral for many language tasks; OpenAI or Anthropic Claude where they meaningfully outperform. The choice is made per-engagement, not per-vendor relationship.
Client data does not leave the client's preferred jurisdiction without written consent. Models trained on client data belong to the client at the end of the engagement, with documented weights and reproduction instructions.
Every model is documented to peer-review-grade. The next consultant — internal or external — can re-train the model from the data and the documentation, without recourse to anyone on this team.
Considering an Auctus AI engagement?
We are taking on a small number of beta engagements through 2026. Pilots are open; build & deploy slots will open from Q3.
Book a 30-minute scoping call →