A golden set, reviewed once.
Then checked by hand.
A healthtech company with a consumer-facing AI feature. Each user fills in a detailed health profile, and the feature returns a personalised set of recommendations drawn from the company's own catalogue. Live, in front of real people making real health decisions.
Where the gaps were
They weren't flying blind. They had a golden dataset — real profiles paired with the recommendations an expert had validated. But that review happened once, externally, and the feedback arrived unstructured. It never became something the team could run against, again and again. Day to day, quality was spot-checked one recommendation at a time, by eye.
It's a familiar shape. Some of these may resonate:
- A golden set reviewed once and never run systematically — only spot-checked, case by case, by a person.
- Expert feedback delivered unstructured, so it never compounded from one iteration to the next.
- No measure of reliability. Run-to-run consistency on the same profile was never measured. Nobody knew whether the feature held up on a second try.
- Quality treated as one broad concept, not named failure modes — and model choice made on cost and latency, never benchmarked on their own data.
We turned a one-off review
into a system.
The golden set they already trusted, run at scale — with every signal compounding in one place. Seven steps:
- 1.Took their validated golden set and made it the foundation — real profiles, each with the expert's expected output, now something they could test against on demand.
- 2.Ran it at scale: the same profiles through multiple models, repeatedly, every run scored automatically against the expected output. Not a spot-check — the whole set, over and over.
- 3.Compounded the signals. Expert judgement, automatic metric checks, and user feedback captured together and carried forward, so each iteration built on the last instead of starting from scratch.
- 4.Turned “quality” into a set of named, automatic checks — schema validity, recommendation count, dosage formatting, expert-aligned selection, catalogue compliance, a safety gate, and output integrity.
- 5.Ran error analysis across every run and every failure — looking at all the misses together, not one at a time.
- 6.Surfaced a systematic bias the team wasn't looking for, and named the distinct failure categories behind it.
- 7.Iterated fast against those named failures — and, because the loop was now systematic, found the model-and-prompt combination that hit their target at the best cost and latency.
What one-by-one checking can't see.
| What we measured | What we found |
|---|---|
| Broad quality score, incumbent setup | 91% · looked healthy |
| Expert-aligned selection, at scale | 50% · the gap one-by-one checks miss |
| Reliability across repeated runs | Unmeasured — scored on every run |
| Named, measurable failure modes | 0 — 6 categories, each with a check behind it |
| Models benchmarked on their data | 1 — multiple, on the same golden set |
| The golden set itself | Reviewed once, by hand — run at scale, on demand |
A pattern no one was looking for.
Running the golden set at scale let us do what a human reviewer can't: look at every failure together. The selection errors weren't random. They were systematic — the model consistently favoured some items and suppressed others, independent of the user in front of it.
| Behaviour across mismatched cases | Frequency |
|---|---|
| One item — systematically omitted, regardless of profile | missing in 14 of 16 |
| A second item — inserted regardless of profile | added in 10 of 16 |
| A third item — inserted regardless of profile | added in 9 of 16 |
A reviewer spot-checking a handful of outputs sees a handful of plausible answers. They have no way to see that one item is dropped on nearly every one, and others pushed in where they don't belong. The pattern only exists across runs — which means manual review, however expert, is structurally blind to it. The team wasn't aware of this bias. They weren't looking for it. It surfaced only because the experimentation was systematic.
When iteration is systematic,
the gains compound.
With a systematic loop in place, improvement stopped being guesswork. The same infrastructure that surfaced the bias also showed that the model they'd chosen — picked on cost and latency, never tested on quality — wasn't the best option on their own data.
| Before | After | |
|---|---|---|
| Model chosen on | Cost & latency, untested on quality | Measured quality on their own data |
| The chosen model | Incumbent frontier model | A smaller, lower-cost, faster model |
| Iteration loop | One-off review, unstructured feedback | Systematic, compounding across runs |
| Expert-aligned selection | 50% baseline | Improved against target with a new prompt |
What they walked away with.
- Their golden set turned into a living test — run at scale, on demand, instead of reviewed once by hand.
- A system where expert judgement, automatic checks, and user feedback compound from one iteration to the next.
- A set of named, automatic failure modes — including the systematic bias one-by-one review can't surface.
- Reliability they can measure: the same set, across models, run after run.
- A model benchmark on their own data, and an improved prompt that moved the metric that matters toward target.
- The ability to catch the next unknown failure before a user does — not after.
If any of these feel familiar.
- You have a golden set or expected outputs, but you only ever check them one by one
- Your expert feedback arrives unstructured and doesn't compound from one iteration to the next
- You can't say how reliable your feature is — whether the same input holds up across runs
- You suspect there's a pattern or a bias in the failures, but no way to see it in aggregate
- You picked your model on cost and latency and never benchmarked quality on your own data
- You're in a regulated or health-adjacent space where “it looked fine” isn't a defensible standard
Then the same engagement applies. We turn your golden set into a system — run at scale across models, with expert insight, automatic checks, and feedback compounding in one place — and surface the patterns one-by-one review can't. You keep all of it.
More results from real projects.
Document extraction — 52% to 88% accuracy in three weeks
A 20-point eval rubric, 8 prompt iterations, and a model benchmark across 4 providers — on the same model family they were already running.
Read the case study →ESG screening agent — 18% to 93% accuracy, 25 experiments
All five top-tier providers failed equally. The production answer turned out to be a cheaper model with a clean rubric and a corrected ground-truth set.
Read the case study →