Classical ML, for when an LLM is the wrong tool
Forecasting, anomaly detection and scoring on your business data, tested against the method you use today.
- Measured against your baseline
- Inside the tools you already use
- Monitored after launch
Problems classical ML solves well
Structured data, a clear number to predict and a decision that depends on it.
Demand forecasting
Forecasts per product and location that feed purchasing and production plans.
Predictive maintenance
Failure risk from sensors and work orders, before a machine stops the line.
Churn prediction
Accounts likely to leave, with the signals behind each score.
Fraud and anomaly detection
Unusual invoices, payments and transactions flagged for a reviewer.
Lead and credit scoring
Rank leads and estimate payment risk from your own history.
Pricing and recommendations
Suggested prices and next-best products, tested before they go live.
What goes in and what comes out
Each model ends inside a screen your team already uses.
| Problem | Typical method | Data it needs | Feeds into |
|---|---|---|---|
| Demand forecasting | Gradient boosting, time series | Sales and stock history | Purchasing plans |
| Predictive maintenance | Survival models, gradient boosting | Sensor and work order data | Maintenance requests |
| Churn prediction | Classification | Orders, tickets, activity | CRM follow-ups |
| Invoice anomalies | Isolation forest, rules | Journal entries, vendor history | A review queue |
| Late payment risk | Gradient boosting | Receivables history | Payment reminders |
From history to a daily decision in five steps
If it can’t beat your current method, it doesn’t ship.
Baseline
Record how today’s method performs, so the target is real.
Features
Build inputs from your data, and fix the errors this exposes.
Backtest
Replay past periods the model never saw, against that baseline.
Deploy
Write predictions into the workflow people already use.
Monitor
Track drift and retrain when performance slips.
Questions we get asked first
When is classical ML better than generative AI?
When the job is predicting a number or a category from structured data, such as demand, failure risk or churn. Models like gradient boosting are usually cheaper to run, easier to test and more consistent for those tasks than a large language model.
How much data do we need?
It depends on the problem and how much the pattern varies. Forecasting benefits from history covering several seasonal cycles, and rare events like failures need enough examples to learn from. A short data audit gives a firmer answer than a general rule.
Can predictions appear inside our existing systems?
Yes. Predictions can be written back to your ERP, CRM or planning tool through their APIs, so a planner sees a suggested quantity or a risk flag in the screen they already use rather than a separate dashboard.
What if the model can’t beat our current method?
Then it shouldn’t ship, and we will tell you. Sometimes the real gain lies in fixing the data or in a simple rule, and that is a useful finding too.
Can machine learning and LLMs work together?
Yes, and it’s often the best design. The ML model makes the prediction and a language model explains it in plain words or lets people ask questions about it, while the number itself stays testable.
Which number do you wish you knew? Let’s predict it.
Name the decision it would change and where the history lives.
Start the conversation