01 — Practice
AI products that are measured, not demonstrated.
A convincing demo and a dependable product are different artefacts. We build the second kind: retrieval that cites its sources, agents whose permissions are explicit, and an evaluation suite that has to pass before anything ships.
Approach
How an engagement actually runs.
Scope against a baseline
Before any model work, we establish how the task is done today and what "correct" means numerically. Without that number, every later claim of improvement is a guess.
Retrieval before fine-tuning
Most failures blamed on the model are retrieval failures. We build the corpus pipeline — chunking, embedding, ranking, citation — and measure it on its own.
Bounded agents
When a system takes actions, each tool gets an explicit permission scope and an audit record. Irreversible operations require confirmation by construction, not by prompt.
Evaluation as a gate
A regression suite runs on every change. Prompt edits are code changes and get reviewed as such. Nothing reaches production on the strength of a demo.
What we build with
The layers we own in a delivery.
- Model access
- Hosted APIs, self-hosted open weights, or both behind one routing layer
- Retrieval
- Vector and hybrid search over private corpora, with citations back to source
- Orchestration
- Tool-calling agents with typed interfaces and bounded permissions
- Evaluation
- Golden sets, LLM-as-judge with human calibration, regression gating in CI
- Interface
- The part people actually touch — designed, not bolted on at the end
- Observability
- Traced calls, token accounting, latency and failure budgets
Where data sensitivity rules out third-party inference, the same architecture runs against self-hosted open-weight models on your own hardware. That decision belongs at design time, not after the fact.
Have a task you think a model could take over?
Send us the workflow and what happens when it goes wrong. If the honest answer is that it is not an AI problem, we will say that instead of building something.