Help

This application demonstrates six AI-powered solutions built for ICL Group, one of the world's leading specialty minerals companies. ICL mines and processes potash, bromine, and phosphates from the Dead Sea and other locations across 13 countries, serving agriculture, food, and industrial markets globally.

Each demo solves a different enterprise problem using modern AI — from answering questions about company documents to forecasting commodity prices and detecting equipment failures before they happen.

Potash is a potassium-rich mineral used primarily as fertilizer. It's one of the three essential nutrients for plant growth (alongside nitrogen and phosphorus). ICL is a top-5 global potash producer, extracting it from the Dead Sea in Israel and from underground mines in Spain and the UK.

Potash prices are influenced by agricultural demand cycles, geopolitical events (the 2022 price spike was driven by sanctions on Belarus and the Russia-Ukraine conflict), weather patterns, and currency fluctuations. This makes price forecasting both valuable and challenging — which is why the Demand Forecasting demo exists.

A conversational AI that answers questions about ICL's operations, financials, products, and strategy — grounded entirely in company documents. Instead of making up answers, it retrieves relevant passages from a knowledge base and cites its sources.

How it works

1. Ingest documents — Click "Load Knowledge Base" to index ICL documents into a vector database using embedding models.
2. Ask a question — Type any question about ICL. The system finds the most relevant document chunks using semantic similarity search.
3. Get a grounded answer — The LLM generates a response based only on the retrieved context, with source citations so you can verify the answer.

Try asking

• What was ICL's revenue in the most recent fiscal year?
• How does ICL extract minerals from the Dead Sea?
• What are ICL's main business segments?

Upload any PDF, invoice, purchase order, or safety data sheet and extract structured data as clean JSON — or chat with the document to ask questions about its content. Works with scanned documents and any language.

Two modes

• Extract Data — Tell it what to extract (line items, payment terms, hazard codes) and get structured JSON output, downloadable as CSV.
• Chat with Document — Ask free-form questions about any uploaded document.

Demo documents included

• Purchase Order — Extract line items, delivery dates, payment terms
• Invoice — Extract shipping details, banking info, HS codes
• Safety Data Sheet — Extract hazard codes, transport classification, chemical properties

Forecast ICL potash prices and sales volumes using two complementary models — then explore bull, base, and bear scenarios for revenue planning.

Controls

• Forecast Horizon — How far into the future to predict (30–365 days). Shorter horizons are more accurate.
• Bull Case Multiplier — Scales the base forecast upward (1.0x–1.5x) for the optimistic scenario. Use higher values to model strong demand or supply disruptions.
• Bear Case Multiplier — Scales the base forecast downward (0.5x–1.0x) for the pessimistic scenario. Use lower values to model demand weakness or oversupply.

Four chart views

1. Price Forecast — Historical prices with the Prophet model's forecast and 95% confidence interval (shaded band).
2. Volume Forecast — Monthly sales volumes (actual vs. predicted) with an average reference line.
3. Scenario Analysis — Bull, base, and bear price trajectories overlaid for comparison.
4. Model Comparison — Side-by-side accuracy metrics for Prophet (time series) vs. LightGBM (gradient boosting), plus a feature importance chart.

Understanding the metrics

• MAE (Mean Absolute Error) — Average dollar amount the prediction is off. Lower is better.
• RMSE (Root Mean Square Error) — Like MAE but penalizes large errors more. Lower is better.
• R² — How much of the price variance the model explains. 1.0 is perfect, 0.0 means no better than guessing the average.
• MAPE — Error as a percentage of the actual price. Under 5% is generally excellent.

Why Prophet usually wins

Potash prices follow strong seasonal patterns (agricultural demand cycles) with a long-term trend. Prophet was purpose-built for this type of time-series decomposition. LightGBM is a general-purpose model that relies on engineered lag/rolling features — powerful, but it doesn't inherently understand time structure with only ~730 training points.

Three specialized AI agents research a topic simultaneously, then an orchestrator synthesizes their findings into an executive intelligence brief.

The agents

1. Market Analyst — Analyzes market conditions, pricing trends, demand drivers, and the short-term outlook for ICL's core markets (potash, bromine, phosphates, specialty fertilizers).
2. Competitive Intelligence — Tracks competitor moves (Nutrien, Mosaic, K+S, Arab Potash, Albemarle), market share dynamics, and ICL's competitive positioning.
3. Regulatory Scanner — Monitors regulatory changes across ICL's 13 operating countries — REACH, environmental regulations, mining permits, trade policies.

How it works

Enter a topic and click "Run Analysis." All three agents start researching in parallel (you'll see their status cards pulse amber). As each agent finishes, its card turns green and you can expand it to read the full report. Once all three complete, the orchestrator combines everything into a structured executive brief with key findings, strategic implications, and recommended actions.

Monitor industrial equipment sensors in real time, detect anomalies before they cause failures, and predict how many days of useful life remain. The demo simulates a potash crystallizer unit at ICL's Dead Sea Works.

Controls

• Monitoring Period — How many days of sensor history to analyze (7–60 days). Longer periods give more context for trends.
• Anomaly Sensitivity — The contamination rate for the Isolation Forest model (2%–20%). Higher values flag more data points as anomalous — use higher sensitivity for critical equipment where false negatives are costly.

Sensors monitored

• Temperature — Operating temp in °C (normal: 65–85, critical: >95)
• Vibration — Bearing vibration in mm/s (normal: 0–4, critical: >6)
• Pressure — System pressure in PSI (normal: 120–160, critical: <100)
• Flow Rate — Material flow in m³/h (normal: 35–50, critical: <25)
• Power Consumption — Draw in kW (normal: 75–95, critical: >110)

Four chart views

1. Sensor Data — Time series for each sensor with anomaly points overlaid and warning/critical threshold lines.
2. Health Score — Overall equipment health (0–100%) over time. Red below 40%, amber below 70%.
3. RUL Prediction — Actual vs. predicted remaining useful life in days, with 30-day and 60-day warning thresholds.
4. Anomaly Detection — Table of detected anomalies and feature importance showing which sensors contribute most to predictions.

Understanding the models

• Isolation Forest — An unsupervised model that identifies anomalies by how easily data points can be "isolated" from the rest. Points that are easy to separate are anomalies. No labels needed.
• Gradient Boosting Regressor — Predicts remaining useful life by learning the relationship between current sensor readings and time until failure. Trained on historical degradation patterns.

Benchmark LLM extraction accuracy on chemical Safety Data Sheets. Compare how different models handle structured extraction from safety-critical documents, with domain-specific validators that catch errors a generic eval would miss.

Two modes

• Eval Mode — Run against a pre-verified SDS document (ICL MKP) with known ground truth. Compares Claude Sonnet, GPT, and Mistral on pass rate, critical failures, and field-by-field accuracy.
• Compare Mode — Upload any SDS PDF. Models extract data side-by-side with disagreement highlighting and hallucination detection (values not found in the source document).

Domain validators

• CAS Number Checksum — Verifies chemical registry numbers using the standard check-digit algorithm. A wrong CAS number could mean the wrong chemical.
• GHS Hazard Codes — Validates that H-codes and P-codes follow the Globally Harmonized System format.
• Source Traceability — Confirms every extracted value actually appears in the source document. Flags hallucinated safety data.

Why this matters

Safety Data Sheets contain information that protects workers, first responders, and the environment. A hallucinated CAS number or wrong hazard classification is not just an error — it's a safety risk. This demo shows how to build AI systems that are verifiably correct for safety-critical domains.