01
Generation and enrichment pipelines
Multi-stage pipelines that produce and enrich content on a schedule, with parameters recorded so any output can be reproduced or explained.
Artificial intelligence
Machine learning that has to survive production.
Model work only counts when it runs continuously, at a known cost, against real inputs, with a person able to explain what it did. Everything here is built to that standard.
- Posture
- Applied, not speculative
- Serving
- GPU infrastructure we operate
- Versioning
- Recorded per model and per run
- Review
- Automated decisions are reviewable
01 / Systems
What we run.
02
Classification and quality gates
Classifiers that decide whether something passes. Precision and recall are measured against held-out sets rather than asserted, and a gate that cannot fail is treated as a defect.
03
Retrieval
Vector retrieval over first-party corpora, serving product features directly rather than sitting in a notebook.
04
Inference serving
Model serving sized to the hardware it actually runs on, with throughput, memory and cost per unit treated as engineering constraints.
02 / Discipline
How we keep it honest.
Every gate gets a negative control. A quality check that has never rejected a known-bad input has not been shown to work, so we test the failure direction as well as the success one.
Nothing is claimed without a measurement. Where a model's accuracy is not established, the system is designed so the model advises and a person or a deterministic rule decides.
Runs are reproducible. Parameters, versions and inputs are recorded together, so an output can be regenerated or accounted for months later.
D1
Versioned models
Every model and prompt revision carries an identifier that appears in the record.
D2
Measured gates
Precision and recall against held-out sets, re-measured when anything changes.
D3
Advisory by default
An unproven classifier informs a decision, it does not make one.
D4
Cost as a constraint
Inference cost per unit is budgeted in the design, not discovered in the bill.
03 / Where it runs
On hardware we specify and operate.
Inference runs on GPU infrastructure we size, configure and maintain, rather than on a managed endpoint whose behaviour and pricing we do not control. That keeps throughput predictable and keeps the data inside systems we operate.
It also means the cost of a feature is something we can state, which matters when the technology is licensed to somebody who has to budget for running it.
- Compute
- GPU, specified and operated in house
- Placement
- Inside infrastructure we control
- Scheduling
- Serialised where hardware is shared
- Observability
- Per-run timing and failure capture
- Recovery
- Pipelines resume rather than restart