Intelligence / Machine Learning & AI
Models in production, not in notebooks.
Applied machine learning and LLM systems built to be deployed, monitored and corrected — not demonstrated once.
Start a projectThe prototype never shipped
A notebook that worked, and no path from there to something a product can call.
It got worse and nobody noticed
Model quality degraded as the inputs drifted, with no monitoring to catch it.
The demo does not survive real users
Impressive on curated examples, unreliable on the messy input people actually send.
What we do
- 01Model deployment and servingInference behind a real API, with latency and cost budgets.
- 02Evaluation harnessesMeasuring quality on your data before and after every change.
- 03Drift and quality monitoringCatching degradation from the metrics rather than from complaints.
- 04Retrieval and LLM systemsRetrieval-augmented systems where the grounding is tested, not assumed.
- 05Human-in-the-loop workflowsReview and correction designed in, for decisions that carry consequences.
- 06Feasibility reviewAn honest answer on whether machine learning is the right tool at all.
How we work
- 01
Frame
1 week
The decision being automated, and the cost of getting it wrong.
- 02
Evaluate
2–4 weeks
A baseline and an evaluation harness before any modelling.
- 03
Build
6–12 weeks
Serving, monitoring and the correction path, together.
- 04
Run
Ongoing
Watched for drift; retrained deliberately rather than on a schedule.
Refurbishment grading support
- The challenge
- Device condition graded by hand, inconsistently between vendors, which directly drove disputes.
- The approach
- An evaluation harness built on historical gradings first, then a model proposing a grade with a human confirming it, and disagreement tracked as the quality signal.
- The result
- Grading proposed automatically and confirmed by a person, with drift visible as disagreement rises.
Questions we get asked
Do we need our own model?
Usually not. A hosted model with good retrieval and evaluation beats a bespoke one for most problems, and costs far less to run.
How do you measure quality?
An evaluation harness on your data, built before the model. Without it, "better" is an opinion.
What about hallucination?
Grounding, retrieval and human review where the decision matters. We will not claim to eliminate it.
Can you work with our data scientists?
Yes. Most of our value here is the engineering around the model rather than the model itself.
Is our data used to train anything?
Not without an explicit written agreement. The default is no.