Machine learning development services that start small
We help small and mid-sized businesses use machine learning where it pays: classifying messages and documents, forecasting demand, search by meaning and evaluating AI systems before they reach customers. We start with your data and a baseline, prefer small efficient models, and put every model behind an API with monitoring. Pricing is scoped after a discovery call.
Use cases
What we build for your business.
Text and document classification
Route emails, tickets or forms to the right queue, with confidence scores and a human check.
Demand and workload forecasting
Predict bookings, orders or call volume from your history to plan staff and stock.
Search by meaning
Embeddings that find the right product, document or answer even when the words differ.
LLM evaluation
Test sets and scoring that catch regressions when prompts, models or content change.
Lead and churn scoring
Rank leads or flag at-risk customers from the data already in your CRM.
Data extraction
Pull fields from invoices, forms and PDFs into structured records you can check.
Trends we track
What's new in 2026 and how we use it.
Small models are competitive
Open models around 3 billion parameters now report results close to larger ones on their authors' benchmarks (SmolLM3 is one), which makes cheaper, private deployments realistic for narrow tasks.
Source: huggingface.coEvaluation is part of the build
Workflow tools now include evaluation: n8n supports quick checks during development and metric-based regression tests on real production cases.
Source: docs.n8n.ioOn-device models
Phones and browsers ship built-in models (Gemini Nano on Android, built-in AI in Chrome), so some predictions can run without sending data to a server.
Source: developer.android.comProject planner
How a project runs, phase by phase.
Machine Learning project plan
Typical time to launch: about 8 weeks, then support
-
01
Discovery
About 1 week
The decision the model supports, the data you have, how it is labeled and what "good enough" means.
-
02
Data and design
1 to 3 weeks
Data audit, cleaning plan, a simple baseline and the evaluation metric agreed up front.
-
03
Build
2 to 4 weeks
Feature pipeline, model training or prompt-based approach, compared against the baseline.
-
04
Test
1 to 2 weeks
Hold-out evaluation, error review with your team, bias and edge-case checks.
-
05
Launch
2 to 4 days
Serve behind an API or inside a workflow, with logging and a human fallback.
-
06
Support
Ongoing
Monitor accuracy and drift, re-train or re-evaluate on a schedule.
Architecture
A typical machine learning pipeline at a glance.
- Input
- Logic and services
- AI step
- People
- Output
How it flows
- 1 Historical data is pulled from your CRM, spreadsheets or logs, cleaned and labeled.
- 2 The feature pipeline turns it into model inputs, and a simple baseline is set before anything clever is tried.
- 3 A small model (or a prompt-based approach on a hosted model) is trained and scored on a held-back test set your team reviews.
- 4 The model is served behind an API or a workflow step, and accuracy and drift are monitored so it can be re-trained when the data changes.
Tech we use
Pricing
Scoped after a discovery call
The cost of a machine learning project depends mostly on your data: how much exists, how clean it is and whether it is labeled. Rather than publish a range we cannot stand behind, we look at your data on a free call and then give you a fixed quote, often for a small proof of concept first.
Book a free 20-minute callOur work
Related case studies from our own products.
We have not published a custom-trained model case study yet. The related work below uses AI classification and evaluation inside real products and workflows.
Order Status & Returns Assistant
A ready-to-import n8n workflow for online stores. AI sorts every customer message, the workflow finds the order in Shopify, replies with tracking or return instructions,...
Read the case studyMake It Flow Platform
makeitflow.si is our own platform and the site you are on. It runs website audits across speed, SEO, security, email and DNS, offers more than 150 free tools, serves indu...
Read the case studyIt depends mostly on your data: how much there is, how clean it is and whether it is labeled. We scope and quote after a free discovery call instead of guessing, and we often recommend a small proof of concept first.
It depends on the task. Simple classification with a few clear categories can work with hundreds of good examples, especially when starting from a pre-trained model; forecasting usually needs at least a year or two of history to capture seasonality.
Often only for a few narrow decisions, such as routing messages or predicting busy days. We check whether a rule, a spreadsheet formula or an off-the-shelf AI model already solves it before proposing a custom model.
For many text tasks a hosted model with good prompts and an evaluation set is enough. A custom or fine-tuned small model makes sense when volume is high, data must stay private or the task is very specific.
We agree a metric and a baseline before building, test on data the model has not seen and review the errors with your team. The model only goes live if it beats the baseline by enough to matter.
Models drift as customers and products change. We log predictions, monitor accuracy against real outcomes and re-train or re-evaluate on a schedule or when performance drops.
Free 20-minute call
Tell us what you want to build
Book a free 20-minute call. We will look at how the work runs today, tell you honestly whether it is worth building and give you a fixed quote if it is.