Consumer AI
Nufeed
A personalised news feed that lets people design their own algorithm instead of inheriting one — taken from spec to a public test-flight in weeks.
<2s
per-topic feed rewrite latency
- Next.js
- LLM pipelines
- Content
Services
AI features that ship
AI in a product, not a demo. The hard part is never the first plausible output — it is knowing whether a change made things better, and keeping the feature usable when the model is slow, wrong or down. This is the work we do most of on our own systems, which is the only reason we are quick at it on yours.
When teams call us
If more than one of these is true, this is usually the right starting point.
Harness in the first fortnight, then the feature. One engineer, occasionally two.
- A prototype produces plausible answers nobody can verify
- You cannot tell whether a prompt change helped or hurt
- The feature works until the model has a bad day
What it covers
- An evaluation harness built before the feature
- Labelled golden sets and per-query scoring
- Regression gates in CI, not in review
- Latency budgets, fallbacks and graceful degradation
- Citations and provenance where the answer matters
What you keep
- The evaluation harness and its corpus
- A retrieval or agent pipeline you can tune
- Model-routing configuration you control
How it runs
How we approach it
The sequence this work follows, and why it is in that order.
Harness before feature
The scoring rig is built first, against a labelled set drawn from your real queries. Without it every prompt change is an opinion, and the team argues from screenshots.
Establish the floor
We measure the simplest thing that could work — often plain retrieval with no model in the loop. Plenty of features never need to beat it, and knowing that early saves the budget.
Iterate against the score
Retrieval, prompting and routing tuned as separate variables, each scored. Changes that do not move the number do not merge, however good the demo looked.
Design for the bad day
Latency budgets, fallbacks, and a defined behaviour when the model is slow, wrong or unavailable. The feature has to degrade into something useful rather than a spinner.
What we build with
The stack for this work
Chosen per deployment, not per fashion. If your team already runs something that works, we work in that instead.
Models
- Claude
- GPT
- Open-weight via vLLM
Retrieval
- pgvector
- Hybrid search
- Rerankers
Evaluation
- Golden sets
- LLM-as-judge
- Human calibration
Serving
- Streaming
- Fallback routing
- Semantic caching
Evidence
Where we have done this
Before you ask
Questions we get about AI features that ship
Also from us
The rest of what we build
Most deployments draw on more than one of these.
- Product engineeringAn engineer deployed to own the whole build — architecture, pipeline and runbooks — working in your accounts from the first week.Learn more
- Web & mobile applicationsFast on a mid-range Android. Accessible to WCAG 2.2 AA. Built on a component library your own team extends without calling us.Learn more
- Platform & cloudInfrastructure your own team can run without a dedicated rotation. Reproducible, deliberately boring, and priced predictably.Learn more
- Legacy modernisationChange code nobody remembers writing. An engineer maps it, pins it with tests, then moves it in steps you can revert one at a time.Learn more
- Reliability & scaleFind out before your customers do. Load models, failure-mode analysis, and an alert catalogue where every entry names an owner and an action.Learn more
Let’s talk
Need AI features that ship?
Talk to an engineer who has done this before — and if this is the wrong place to start, they will say so.