SpamShield email classifier
MAUK R&D · applied machine learning
A spam classifier built from first principles that explains every verdict it gives.
Paste an email and get a spam or legitimate verdict in under a millisecond, together with the words, symbols and capital-letter patterns that drove it. It was trained on the public 4,601-email UCI Spambase dataset, and the app publishes its own accuracy and mistakes instead of asking to be trusted.
- What it is
- AI system
- Industry
- Cross-industry
- Status
- Live
- Year
- 2026
The problem
What needed solving
A classifier that only outputs 'spam' is hard to trust and impossible to debug. This build had to show its working: which signals fired, how confident it was, and how often it is wrong, including the mistake that matters most, a real email marked as spam.
The solution
What we built
An email inspector with a live report on the model beside it, written from scratch so every step of every verdict can be explained.
Core functionality
- Verdict with a confidence figure and risk level
- The trigger words, symbols and capital-letter patterns behind each verdict
- Four sample emails for quick testing
- A model report with its accuracy and every kind of mistake it makes
- A retrain button that reshuffles the data and re-checks the results
- An interface other systems can call to check emails
Screens
More screenshots
Captured from the live site, a demo workspace, or the product running on sample data. No customer records, passwords or private admin data are shown.
Business impact
What it achieved
83.8%
accuracy on 920 emails the model never saw
Trained on 3,681 emails from the 4,601-record UCI Spambase dataset.
Verified96%
of spam caught
14 spam emails missed out of 369 in the test set.
Verified0.9 ms
to classify an email
Measured in the app on a pasted 52-word email.
Verified10 / 10
machine-learning tests passing
Run on a clean install.
Verified
What sets it apart
Why it works
- 01
Explains itself
Every verdict shows the exact words and patterns that drove it, so people can see why an email was flagged.
- 02
Transparent about accuracy
The app publishes its own score and its errors, rather than one flattering number.
- 03
Honest about the ceiling
It documents why this approach levels off near 84% on this data, and what would push it higher.
- 04
Tested
Its maths, features and scoring are covered by automated tests that pass on a clean install.
Services
Where this work fits
Related projects
Similar projects
LiveIn-house buildVerifiedMulti-disease risk screening model
MAUK R&D · applied machine learning
Risk scores for ten conditions from 25 routine measurements, trained on 281,000 records and private enough to run on the patient's own device.
86.88%accuracy on patient records the model never saw
Machine learning
View project
In-house buildVerifiedExam score prediction system
MAUK R&D · applied machine learning
From a student dataset to a working predictor: two models compared honestly, and the one that won shipped.
0.617fit on students the model never saw (R²)
Machine learning
View project
ResearchVerifiedRoad-scene object detection on BDD100K
MAUK R&D · computer vision
A road-scene detection model trained from scratch on 70,000 images with a laptop GPU, and an honest account of where it struggles.
0.575detection accuracy (mAP50) on 10,000 unseen images
Machine learning
View project
Discuss a similar project
Building something like SpamShield?
We'll tell you what we'd reuse, what we'd change and what it would take.