0%
Less manual monitoring
Dropout risk agent, replacing termly review
Applied AI · London, United Kingdom
We build AI systems that
We build AI systems that
Twelvetech Systems is an AI software company. We build the applications, train the models, run the infrastructure and get you found, for organisations that have to justify the decisions their software makes.
0systems shipped
0UK contracts signed
0services offered
2025incorporated in England
decision tracelive
signal inattribution backevidence
London Brookes CollegeNavv Associates LtdEast London TutorsRegent College LondonOxford AIEOUIEEE Standards AssociationUniversity of Greater ManchesterPearson BTECLondon Brookes CollegeNavv Associates LtdEast London TutorsRegent College LondonOxford AIEOUIEEE Standards AssociationUniversity of Greater ManchesterPearson BTEC
What we do
Six practices. One standard.
If it informs a decision about a person, it has to be inspectable. Hover any tile to see it come alive — each one animates the kind of work it describes.
010203040506
AI and machine learning
Custom models trained on your data — prediction, computer vision, LLM agents and retrieval — evaluated against a cheap baseline before anything sophisticated gets built.
PyTorchTensorFlowRAGMask R-CNNOllama
Software development
Custom applications, multi-tenant SaaS, mobile and integration.
ReactFastAPIPostgreSQL
Growth and SEO
Technical audits, structured data, Core Web Vitals and publishing automation.
Technical SEOSchemaAPIs
Cloud and support
Pipelines, monitoring, migrations and managed care after launch.
AzureDockerCI/CD
IoT and edge
Sensor rigs, firmware and on-device inference where the signal is unreliable.
ESP8266TF LiteFirebase
Explainability, built in rather than bolted on
Where a model touches a decision about a person, the explanation and the route to contest it are part of the same deliverable. Audit trails, tenant isolation and transparency portals are quoted with the build, not sold back to you afterwards.
AttributionAudit trailsUK GDPRContestabilityModel cards
Try it yourself
Most AI gives you a number. Ours shows the reason.
This is a working miniature of what ClassAura does. Switch the input signals on and off and watch the network re-run: the risk score moves, and the cyan attribution pass traces the result back to whichever input actually drove it.
Learner signals
Turn these on and off. Every change re-runs the model in real time.
A demonstration built on the same idea as the production system, with illustrative weights rather than a live cohort.
flag raised
0.00risk score
Outcomes
Four numbers, with the context attached.
A percentage with no context is marketing. These are improvements measured or estimated at delivery, in the setting named under each one.
0%
Satisfaction uplift
Task Tracker Hub, after the interface rebuild
0%
Fewer reporting errors
Asset management, procurement to disposal
0%
Records preserved
Domain migration across Microsoft 365 and Moodle
Products
Nine products. Four worth your time first.
Switch between them to see what each one actually shows a user. ClassAura is live with paying customers; the others range from active development to proven engines ready to build on.
In production
ClassAura
Engagement intelligence for higher education
Reads the engagement signals a college already produces — attendance, participation, submissions, tutorial contact — and returns a ranked, explainable view of which learners need attention this week, with a student portal so nobody is scored in the dark.
- React 18, TypeScript and Vite; FastAPI, SQLAlchemy and Pydantic v2 behind it
- Multi-tenant isolation across thirteen core tables on an organisation_id pattern
- Microsoft Graph integration for Teams and Zoom signals, Entra ID for identity
classaura · cohort view
In development
ClassAura Mark
AI marking assistant with an evidence trail
Drafts formative feedback against your own assessment criteria, and links every sentence it writes back to the passage and criterion that produced it, so a marker accepts, edits or rejects on sight rather than trusting a black box.
- Retrieval-grounded generation against your own briefs and grading criteria
- Every comment traceable to the evidence and criterion behind it
- Marker stays the decision-maker; the assistant never releases a grade
mark · script 04
Available to build
FieldSense
Sensing, on-device decisions and remote action
The monitoring platform behind our agriculture and livestock work, generalised: a sensor rig, a model that runs on the handset, and remote control of whatever needs to happen next. Built for places with poor connectivity.
- ESP8266 sensor rigs streaming moisture, temperature, humidity and level data in real time
- On-device inference so a decision does not wait for a network round trip
- Remote actuation — pumps, relays, alerts — closing the loop from reading to action
fieldsense · plot 3
Available to build
AutoInspect
Vehicle damage assessment from a photograph
Pixel-level detection of scratches and dents, returning where the damage is and what kind it is. Built for claims, fleet checks and rental handovers that currently rely on someone walking around the car with a clipboard.
- Instance segmentation with Mask R-CNN, localising damage rather than classifying whole images
- Separates scratch from dent at mask level, with confidence per region
- Delivered as an API so it drops into an existing claims or handover flow
autoinspect · claim 8821
How we work
Six phases. Scroll to walk through them.
Most AI projects fail before anyone writes model code, because the wrong decision was targeted or the data never supported it. Both are discoverable in the first fortnight if you look.
Phase 01
Frame
Before a model, the decision.
Phase 01 · Week 1
Frame
Before a model, the decision. Who acts on the output, how often, on what information, and what a wrong call actually costs.
- Decision map and success measures
- Stakeholder and approval chain
- The explicit list of what we are not doing
Phase 02 · Weeks 1–2
Audit
What data exists, what is missing, what is biased and what cannot lawfully be used. Most projects narrow here, and that is the point.
- Data inventory and quality report
- Legal basis and retention assessment
- Revised scope, usually smaller
Phase 03 · Weeks 2–3
Baseline
A simple, cheap approach gets built and measured before anything sophisticated. If the baseline is good enough, we tell you and stop.
- Working baseline with measured performance
- The number every later version must beat
- Go or no-go recommendation
Phase 04 · Weeks 3+
Build
Design and engineering in reviewable increments, with the explanation built alongside the model rather than bolted on at the end.
- Fortnightly demos of working software
- Test suite and CI pipeline
- Architecture and API documentation
Phase 05 · Before launch
Harden
Access control, audit logging, monitoring, accessibility, security review of every endpoint, and a defined route for a person to challenge a decision.
- Security and penetration-style review
- WCAG conformance pass
- Runbook and incident procedure
Phase 06 · After launch
Run
Into production with monitoring, drift checks and a support window. Handover means your team can operate it without calling us.
- Production deployment
- Monitoring and alerting
- Handover sessions and 30-day support
Selected work
Twenty-two systems, and what each one changed.
Every card carries its own generative visual. Open one for the case note: what the system does, how it was built and what it changed.
ClassAura · Twelvetech Systems Limited
ClassAura engagement intelligence
Teams-native engagement intelligence for UK higher education, in production with paying colleges.
Open case noteLearning analytics · London
Dropout risk prediction agent
An agent that surfaces at-risk learners before the risk becomes an outcome.
Open case noteTwelvetech Systems Limited
AI desktop assistant agent
A Windows assistant that runs local or hosted models and takes the repetitive edges off a working day.
Open case noteApplied NLP · London
Mental health language classification
A context-aware platform that reads written input for behavioural indicators without overreaching into diagnosis.
Open case noteComputer vision · CarDD dataset
Scratch and dent detection in vehicles
Instance segmentation that finds surface damage on vehicle photographs and says exactly where it is.
Open case noteDeep learning · Flickr8k
Image captioning with attention
An encoder–decoder model that writes a sentence for a photograph, attending to regions as each word forms.
Open case noteWhy clients pick us
Six moments where a supplier's choices become visible.
Every column on the right describes something we have watched happen to a client before they came to us. Worth asking whoever else you are talking to.
| The moment | Twelvetech | What often happens instead |
|---|---|---|
| Who answers your enquiry | One of the two founders, who will also work on the project | An account manager, then a handover to people you have not met |
| What happens in week two | A data audit that often shrinks the scope | A kick-off deck and a signed statement of work |
| When a model is the wrong answer | We say so, and the discovery fee stands | The model gets built anyway |
| Who owns the output | You do, on payment, in writing | Ambiguous, resolved in the vendor's favour |
| What handover means | Docs, tests and sessions until your team can run it | A credentials handover and a support quote |
| How you leave | Thirty days' notice, documentation written for a successor | Difficult, by design |
In their words
Three things clients have said.
Attributed by role rather than name, because most of our clients are institutions with their own communications approvals.
The thing that convinced us was being told, in week two, that half of what we had asked for was not supported by the data we held. Nobody had said that before.
Head of QualityUK further education provider
Our data protection officer had a list of eleven questions. They answered all of them on the first call, including the two we expected them to dodge.
Director of ITLondon college
It went live during term. Staff noticed the new domain and did not notice anything else, which was exactly the brief.
Operations ManagerHigher education provider
Go deeper
Six pages for the questions buyers actually ask.
Technical evaluators and business decision-makers want different evidence, so each has its own path rather than one compressed page.
For the buyer
Case studies
Three production deployments in full: the problem as the client stated it, what we did, and how the outcome landed.
For the buyerIndustries
Eight sectors, what we build in each, and the specific evidence behind every claim.
For the buyerPricing
Starting figures on the page, what moves the number, and three practices we have chosen against.
For the evaluatorTechnology
Stack, tenant isolation, audit logging, data residency and what happens to a model after launch.
For the evaluatorHow we work
Six delivery phases with the deliverables you receive at the end of each one.
For everyoneFull FAQ
Sixteen questions on commercial terms, data handling, model drift, accessibility and project failure.
Bring us the decision you cannot currently justify.
Tell us what you are trying to improve and what data you hold. If a model is not the right answer, we will say so before you spend anything on building one.