A baseline first
The simplest model that could work, measured honestly. If it is good enough, we tell you and stop. You are not paying for complexity that earns nothing.
Data Science & ML
Models that reach production, with the monitoring to keep them honest.
Most models die in a notebook. We build the ones that ship, served behind an API, measured against a baseline, and watched for the day the data shifts underneath them.
What you get
An evaluation set
A held-out set and the metric that matters to your business, agreed before training. Every later change is measured against it.
Served predictions
Batch scoring or a real-time endpoint in your cloud, versioned, with the training code in your repository.
Retrieval that cites
For document and LLM work: chunking, retrieval and prompts you can inspect, with answers traced back to source passages.
Drift and cost monitoring
Alerts when input distributions move, accuracy decays or token spend jumps, so you hear about it before a stakeholder does.
A retraining runbook
Written instructions for when and how to retrain, so the model outlives the engagement and does not quietly rot.
Stack we use
Python · PyTorch · scikit-learn · XGBoost · MLflow · LangChain · pgvector · Claude API · FastAPI · Airflow · AWS · GCP
Engagement shapes
How this work is usually bought
Proof of value
Fixed scope, ends in a measured yes or no.
Embedded data scientist
Monthly, working against your data in your cloud.
Model retainer
We watch drift and retrain on an agreed cadence.
Tell us what you are trying to predict
Describe the decision the model would inform and what data you already hold. We will tell you whether it is a modelling problem or a reporting one.
or email hello@siyon.co.nz