How do you run a spike when you do not yet have the data you would need in production?

AgroFleet Exchange is a marketplace where farmers buy, sell, and rent used tractors, sprayers, and other equipment. FieldScout is the add-on that lets a farmer photograph a leaf and get a quick read on whether it needs spraying. Kwame Delacroix-Nkomo is the AI PM who had to spike it before a single real farm photo existed.

How to answer this, stage by stage

Nobody is scoring whether you found a clever workaround for missing data. They're scoring whether the spike itself can tell the difference between a case it has seen and one it hasn't.

Let's learn

The tablet is bolted to the tractor's dash with two zip ties, cracked screen protector, dust in every seam.

Before FieldScout, a farmer walked the rows looking for wilting or spotting by eye, and when something looked wrong, drove into town to ask the local agronomist, losing most of a day. With FieldScout, a photo gets a read in under three seconds, and the farmer decides on the spot whether to spray or wait.

The second step is the one FieldScout tried to shortcut, before it had ever seen a real leaf.

Here's the turn: no real farm photo existed when the spike had to run, so FieldScout trained on public, professionally lit crop-disease photos instead, and hit 91 percent against that proxy set. The real question was never whether that number was good. It was whether the spike would know the difference once a real, messy farm photo showed up.

The one design decision everything else hangs on: it never claims a diagnosis, only a confidence band and an honest flag.

At its worst, a farmer trusts a confident-sounding answer the model was never actually tested on, sprays for the wrong thing, and loses a week and a chemical bill on a blight the tool had never once seen a real photo of.

What I would leave alone: for the one or two most common conditions, well represented in the public proxy data and easy to spot even under bad lighting, the spike's confident reads held up fine against real farm photos from the very first week.

The lesson: a spike without real data is not a smaller version of the real product. It is the actual process of getting the real data, disguised as a demo, if you design it that way on purpose.

Now here is the same thing as a story

The short version above is what you'd say defending the design in a planning review. Read this one for what it felt like the week a new hire's question exposed what the tablet had never actually been tested against.

Every Thursday, Kwame pulled the week's FieldScout photos and skimmed them for anything odd before the Monday product review.

Same words, photo of a leaf, describing two very different kinds of image.

The spike launched well. Farmers who tried it liked the speed, and for the first few weeks the confidence bands mostly read high, green, trustworthy.

A new agronomy intern, reviewing a batch of low-confidence photos, asked Kwame plainly: "Why does it do great on the demo photos but barely says anything useful on real farmers' own shots?"

Week four is where the gap between proxy data and real data first became visible.

Looking closer, Kwame found real farmers had quietly learned to photograph a single leaf, centered, held against a shirt for a plain background, since that framing reliably got a confident read. Whole-plant shots and natural, cluttered backgrounds, what a farmer would show an agronomist in person, scored low far more often.

The real question was never whether farmers were framing photos "wrong." It was whether the spike's own design had quietly taught them what the model could and couldn't handle, without anyone meaning to.

The anchor's whole job: even on a photo the model has never seen, the farmer never gets a confident wrong answer.

When the confidence band was first designed, someone said, "let's keep the threshold generous so it feels useful," and it sounded reasonable, since at the time nobody knew real photos would look so different from the proxy set.

Naming what waits for real data is what kept the spike honest about what it actually was.

Rerun the same ten weeks with the local photo set already growing from day one, exactly as the anchor intended: by week ten, four hundred real farm photos have been banked through the low-confidence flow, and real-batch accuracy climbs from 63 to 84 percent, without anyone waiting on a second data-collection project that was never going to get funded on its own.

What I'd tell myself, hearing that intern's question: the spike was never really about hitting 91 percent on photos we already had. It was about building the pile of photos we didn't.

SPARK, the plan for a spike with no data yetNot a script for pretending a proxy dataset is the real thing. SPARK is what makes the spike say so, out loud, every time it isn't sure.

The recap, one line per letter: situation is a farmer's day without FieldScout, payoff is trusting a quick read for the obvious cases, anchor is the confidence band plus the "no local data yet" flag plus saving every real photo, risk is an unfamiliar case routing to a person instead of a confident guess, and keep out is everything that waits on real local data.

And if you want to be sure it really works, try it somewhere elseSame five letters, a library consortium instead of a farm. The proxy data changes shape entirely, the honest flag doesn't.

Wilhelmina Osei-Bonsu runs product at Fenwick County Library Consortium, where CatalogMind drafts subject tags for new books before a human cataloger confirms them. No real patron search data existed yet when the spike ran, so the team trained CatalogMind on public library metadata records instead. Mapped onto SPARK: situation is a cataloger manually assigning tags from a printed style guide, one book at a time. Payoff is catalogers reviewing drafted tags instead of writing every one from scratch. Anchor is showing which tags CatalogMind is confident about versus guessing, with every guess reviewed before it ever reaches a public search. Risk is a tag drafted for a book type the public records never covered, like a local self-published history, getting flagged low-confidence instead of published silently wrong. Keep out is fully automatic tagging for anything outside the consortium's most common categories.

A different flip than FieldScout's: once a few drafted tags looked embarrassing, catalogers quietly stopped marking which ones were AI-drafted at all.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "run it on the closest honest proxy, and make it say when it's guessing," and stop.
Cost: there's no budget yet to collect real data separately. Say so honestly, and design the spike itself to collect that data as a byproduct, rather than treating data collection as a future, unfunded project.
The proxy data turns out to be closer than expected, for real: if real farm photos match the stock set better than feared, that's still worth confirming against the confidence bands, not assumed safe just because the headline accuracy looked fine.

Where people run it wrong.
They treat a spike's proxy-data accuracy as if it already describes real production conditions.
They let a model answer confidently on data it has never actually seen, with nothing catching the gap.
They wait for a second, separate project to collect real data instead of designing the spike to collect it from day one.

How to use it live. The moment someone asks how to spike without production data, ask yourself: what's the closest honest proxy, and how will the spike admit the moment a real case looks nothing like it. Name both, and the rest of the plan follows.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0