CapabilityAI & Models

Models
that ship.

A demo that works once is not a system. We build the training, retrieval, evaluation, and serving layers that keep a model useful after launch.

Service 01

Adapt

Fine-tuning and continued pre-training on open-weight models, so domain knowledge lives in the weights rather than in an ever-growing prompt.

  • Supervised fine-tuning
  • Preference optimisation
  • Continued pre-training
Service 02

Ground

Retrieval pipelines that keep answers anchored to your corpus — with the ingestion, chunking, and refresh problems treated as first-class engineering.

  • Ingestion and indexing
  • Hybrid retrieval
  • Freshness and permissions
Service 03

Serve

Inference deployed behind an API you can put in front of users, with batching, caching, and autoscaling tuned to the latency target.

  • Throughput and latency tuning
  • Autoscaling and failover
  • Tracing and cost per request
Method

Evaluation
is the product.

Without a measurement you trust, every model change is a guess. We build the harness before we tune anything.

Your evaluation set, not a public benchmark

Leaderboard scores say little about your domain. We build a task set from your real traffic and failure cases, and hold every change against it.

Open weights where they win

Frontier APIs and open-weight models each have a place. The choice is made on cost, latency, and data residency — then revisited as both move.

Cost per request as a first-class metric

Quality that is unaffordable at volume is not quality. Token economics are tracked alongside accuracy from the first prototype.

Guardrails and traceability

Inputs, outputs, and retrieved context are traceable, so regressions can be diagnosed rather than argued about.

Lifecycle

Prototype to production.

01

Baseline

Define the task, assemble the evaluation set, and measure an off-the-shelf model honestly.

02

Improve

Retrieval, prompting, fine-tuning — applied in the order that buys the most against the measured gap.

03

Harden

Latency, throughput, failover, and guardrails tuned against production traffic shapes.

04

Iterate

Live telemetry feeds the evaluation set, so the system improves with use instead of drifting.

Serving

What runs in production.

Specified per engagement, sized against your latency and throughput targets.

Model hosting
Open-weight models served on Luno compute, or routed to frontier APIs where that is the better trade.
Latency shaping
Continuous batching, KV-cache reuse, quantisation, and speculative decoding where the accuracy budget allows.
Retrieval
Hybrid dense and lexical retrieval with re-ranking, permission filtering, and scheduled index refresh.
Evaluation
Versioned task sets, regression gates in CI, and human review queues for the cases automation cannot judge.
Observability
Request tracing, token accounting, quality sampling, and per-feature cost attribution.
Data handling
Deployment inside your boundary where residency requires it, including fully isolated environments.

Bring us the task.

Describe what the model has to do and how you will know it worked. We will tell you what it takes to get there — and what it will cost to keep it there.