Machine learning built on your data, measured honestly
Gjethe builds machine learning models for business problems: demand and sales forecasting, classification and scoring, recommendations, anomaly and fraud detection, text classification and extraction, and churn prediction. We start by checking your data and a simple baseline, evaluate models on data they haven't seen, deploy them behind an API or into your systems, and monitor them so they keep working as the world changes.
- Free first consultation
- Written quote for your project
- Flexible payment milestones
At a glance
- Problems
- Forecasting, classification, recommendations, anomalies
- Data
- Tables, text, time series, events
- Stack
- Python, scikit-learn, XGBoost, PyTorch
- Deployment
- APIs, batch jobs, inside your apps
Definition
What is machine learning?
Machine learning is a set of methods that learn patterns from historical data to make predictions or decisions about new data, such as forecasting demand, flagging unusual transactions or recommending products.
Machine learning can be valuable when a decision is made often, depends on patterns in data, and is currently made by rules of thumb or not at all. It can also absorb a lot of money with little to show. The difference usually comes down to three things: a clear decision the model supports, enough relevant data, and honest evaluation against a simple baseline before anything is deployed.
Who is it for?
Retail and e-commerce
Demand forecasts, product recommendations and return predictions.
Fintech and finance teams
Risk scoring, anomaly detection and transaction categorisation.
SaaS companies
Churn prediction, lead scoring and usage-based insights.
Logistics and operations
Delivery time estimates, capacity planning and maintenance prediction.
What's included, and what you keep
What's included
- Problem framing and success metrics
- Data audit and feasibility check
- Baseline models and rules for comparison
- Feature engineering
- Model training and tuning
- Evaluation on held-out data
- Explainability reports
- Deployment as an API or batch job
- Monitoring for data drift and accuracy
- Retraining pipeline
- Documentation and model cards
What you receive
Handed over at the end. Yours to keep.
- A feasibility report with baseline results
- Trained models and training code
- Deployed prediction service or batch job
- Monitoring dashboards and alerts
- Model documentation describing data, performance and limits
Start with the decision, then the data
Every useful model supports a decision: how much stock to order, which customers to contact, which transactions to review. We define that decision, how it's made today and what a better decision is worth before choosing any technique. That makes it possible to say later whether the model helped.
Then we check the data. Is there enough history? Does it contain the signal? Are labels reliable? A short feasibility phase answers these questions cheaply and avoids months of work on a problem the data can't support.
Beat a simple baseline first
A complex model is only worth it if it clearly beats something simple, like last year's sales for the same week or a rule your team already uses. We always build that baseline and compare against it on data the model hasn't seen. Sometimes the simple method wins, and we'll tell you.
Classic machine learning or a large language model?
| Problem | Usually best |
|---|---|
| Forecasting numbers from history | Classic time-series or gradient-boosting models |
| Scoring or classifying structured records | Gradient boosting such as XGBoost or LightGBM |
| Understanding or generating text | Large language models, sometimes fine-tuned |
| Answering questions from documents | Retrieval-augmented generation |
| Spotting unusual events | Anomaly detection models plus rules |
Models degrade, so we monitor them
Customer behaviour, prices and processes change, and a model trained on last year's data slowly becomes wrong. We monitor input data and prediction quality, alert when they drift, and set up a retraining process so the model stays useful.
Fairness, privacy and explainability
Models that affect people, such as credit, hiring or pricing, need extra care. We check for unequal performance across groups, use only data you're permitted to use for the purpose, and provide explanations of which factors drive predictions. Some uses are regulated, and your legal advisers confirm the obligations that apply.
Fitting into your systems
A model that lives in a notebook helps nobody. We deploy predictions where decisions happen: as an API your app calls, a nightly job that writes scores to your CRM, or a dashboard your planners use. The integration is planned from the start so the model's output arrives in a form people can act on.
Signs machine learning could help
- A decision is made often, at scale
- You have historical data about past outcomes
- Current decisions rely on gut feeling or fixed rules
- Small improvements are worth real money
- Someone will act on the predictions
How machine learning works, step by step
Every step has a clear output you can review. Timing is agreed in your quote.
- 1
Frame
Decision, success metric and constraints agreed.
- 2
Feasibility
Data audit and baseline results.
- 3
Model
Features, training and evaluation on unseen data.
- 4
Deploy
API, batch job or integration into your tools.
- 5
Monitor
Drift and accuracy tracking with alerts.
- 6
Improve
Retraining and iteration based on results.
Tools, and what shapes your quote
Pricing is based on your project and agreed in your quote. These are the things that move it.
| Factor | Why it matters |
|---|---|
| Data readiness | Clean, accessible data shortens the work; scattered data needs engineering first. |
| Problem complexity | Some problems need several models or custom approaches. |
| Accuracy requirements | Higher stakes need more evaluation and safeguards. |
| Deployment | Real-time APIs need more engineering than batch scores. |
| Monitoring and retraining | Ongoing operation adds infrastructure and support. |
| Regulation | Regulated uses need extra documentation and review. |
| Area | Tools |
|---|---|
| Modelling | Python, scikit-learn, XGBoost, LightGBM, PyTorch, statsmodels |
| Data | pandas, Polars, dbt, BigQuery, PostgreSQL |
| Deployment | FastAPI, Docker, AWS SageMaker, Vertex AI, Modal |
| Tracking and monitoring | MLflow, Evidently, Weights & Biases |
Common mistakes to avoid
What we see go wrong most often, and how a good project avoids it.
Starting with the algorithm
Choose the decision and metric first; the technique follows.
Testing on training data
Models always look good on data they've seen. Evaluate on held-out data.
No baseline
Without a simple comparison you can't tell whether the model helps.
Deploy and forget
Models drift. Monitor and retrain.
Frequently asked questions
How much data do we need?
It depends on the problem. A short feasibility phase looks at your data and tells you honestly whether it's enough.
Can you guarantee accuracy?
No one can before seeing the data. We report expected performance on held-out data and compare it with a simple baseline.
Do you use ChatGPT or similar models?
For text problems, large language models are often the right tool. For numbers and structured data, classic models are usually better and cheaper.
Can the model explain its predictions?
Yes, to a useful degree. We provide explanations of which factors drive predictions overall and for individual cases.
Where does the model run?
In your cloud account or ours by agreement, as an API, a scheduled job or inside an existing system.
What happens when the data changes?
Monitoring detects drift and a retraining process updates the model, with checks before a new version goes live.
Can you work with our data team?
Yes. We can lead the project or support your team with specific modelling, deployment or monitoring work.
Who owns the model?
You do, along with the training code and documentation, once the project is paid.
Do you build models for regulated decisions?
We can, with extra documentation, fairness checks and human review. Your legal advisers confirm the specific obligations.
How is machine learning development priced?
It is quoted per phase, starting with a feasibility study, so you only commit to full development once the data looks promising.
Can you improve a model we already have?
Yes. We review the data, features and evaluation, compare against a baseline and suggest targeted improvements.
How do we know the model is worth deploying?
We compare it with the baseline on held-out data and estimate the value of better decisions, so you can decide with numbers.
Can you build a recommendation system for our shop?
Yes, from simple 'often bought together' rules to personalised models, tested against your current recommendations.
Related services
Data Engineering
Pipelines and warehouses that keep your data reliable.
Computer Vision
Image and video recognition, inspection and OCR.
RAG Systems
AI that searches your documents and cites them.
Data Analytics
Clean data, clear analysis and decisions you can defend.
AI API Integration
Language, vision and speech models in your software.
Machine Learning by industry
Find out what your data can predict.
Tell us what you need. You get an honest assessment, a clear plan and a written quote. The first consultation is always free.
- What happens next
- 1You tell us what you need
- 2We aim to reply within 24 hours with questions
- 3You get a written quote for your project