ML systems, end to end
Computer vision, forecasting and anomaly detection taken past the notebook: trained, evaluated, containerised, deployed and monitored.
- PyTorch
- scikit-learn
- MLOps
- Model serving
Machine learning & data engineering
Models are the easy part.
Propagated Gradients designs, ships and operates production ML, LLM and real-time data systems: the ingestion, the evaluation, the serving and the platform underneath. Then we hand it to your team, running.
Shipped inBankingIoTRetailEnergy
Capabilities
Four things, done properly.
A small practice with a deliberately narrow surface. If a job needs something outside this, we will tell you so rather than learn it on your budget.
LLM and RAG platforms
Retrieval pipelines that cite their sources and stay cheap. Embeddings, vector search, evaluation harnesses and the ingestion plumbing behind them.
- RAG
- Embeddings
- Vector search
- Evaluation
Real-time data engineering
Streaming platforms and lakehouses that hold up under load: Kafka and Flink topologies, Iceberg tables and ETL that does not need babysitting.
- Kafka
- Apache Flink
- Iceberg
- Lakehouse
Platform and enablement
Kubernetes, CI/CD and the deployment path around a model. Plus the hands-on work of leaving a team able to run it without us.
- Kubernetes
- CI/CD
- AWS / GCP
- Team enablement
Live demo
The name, running in your browser.
A real network (1 → 8 → 8 → 1, tanh, Adam) training on this page right now. Every edge is drawn from its actual weight: cyan positive, violet negative, thickness by magnitude. Nothing here is a recording.
A small neural network, trained live.
The interactive version needs JavaScript. It trains a three-layer perceptron to fit a sine, a step or a bump, drawing the weights as they move and the loss as it falls.
The forward pass carries the input to a prediction. The error is measured, and the gradient of that error is propagated back along every edge, each weight nudged by exactly its share of the blame. Repeat a few thousand times and the curve fits. It is also a fair description of how we run an engagement.
Method
One loop, run until it converges.
Every engagement is the same four stages. They are named after the training loop because that is the shape of the work: measure, correct, repeat, stop when it is good enough.
Forward pass
Understand the signal
A short paid discovery: your data, your constraints, what "working" would mean in numbers. You get a written plan whether or not we build it.
Loss
Agree the metric first
We settle on the metric before the model. An evaluation harness and a baseline come first, so every later claim can be checked.
Backward pass
Ship and correct
Short iterations against that metric, in your environment, behind your CI. Each one ends with working software in production.
Converge
Hand over
Documentation, runbooks and pairing until your team owns it. The engagement ends when you no longer need us.
Track record
Numbers from systems that are still running.
Every figure below was measured on a system our founder delivered, in banking, IoT or retail.
- Transactions a day
- 7M+scored by a rules engine built for credit and fraud decisioning
- Records a minute
- 1M+through a near real-time lakehouse and stream-processing platform
- Faster to deploy
- 75%requirement to production cut from months to weeks by CI/CD and agile process
- In production ML
- 7+ yrsacross banking, IoT and retail, including leading a team of ten
Get in touch
Tell us what is not working.
A model that will not leave the notebook, a pipeline that falls over at 3am, a RAG system that invents things, or a team that needs someone alongside them for a quarter. Describe it in a paragraph and you will get a straight answer about whether we are the right people, including when we are not.