How we unblocked a $58M first-year platform opportunity with production AI in under 6 weeks.

Their product couldn't ship with AI that wasn't ready.

The company built drug-discovery technology for pharmaceutical research. They wanted to place their platform in partner labs, where researchers could use it to identify promising antibody candidates.

The core AI had to review images and select candidate cells. Manual review relied on experienced scientists. The work took too long to keep up with the volume of images, and the team couldn't inspect everything before biological time limits forced a decision.

They had already rented a computer-vision service at ~$200K a year. When it fell short, they tried building internally. Five months and $590K in engineering resources later, that effort still hadn't reached production.

With both attempts falling short, they came to Already Built to replace the computer-vision backend. Their goal was clear: get the AI ready for a platform launch with $58M in first-year potential.

Two researchers could label the same image differently.

Two different approaches had failed, so we examined what both were learning from: the cases their team said the vendor had classified incorrectly. Before training another model, we needed to know whether those reference labels were reliable.

In a two-hour session, three of their researchers reviewed ~50 cited errors after we taught the labeling approach. ~60% turned out to be label errors. The other 40% were genuine misses by the tool.

That led us to compare labels their own reviewers had assigned to the same images. They disagreed because no one had written a clear, shared definition of a valid cell.

Both the vendor tool and the internal build had been learning from that inconsistent reference. Replacing the model alone would leave the same problem in place.

We wrote down what a good answer looked like.

Which is why our first deliverable was annotation guidance, before another model. It defined a valid cell, showed accepted and rejected examples side by side, and gave reviewers an explicit way to mark cases they couldn't determine.

Abstract instructions still left room for disagreement on difficult images, so we built a visual reference their experts could use to settle those cases and train annotators.

That changed the division of work. Their researchers set the standard and resolved difficult cases. Trained annotators could apply it to the larger dataset, so the relabeling no longer depended on scientists doing every repetitive step.

We also brought the imagery back onto their own hardware and rebuilt the working dataset from the original instrument files. The company regained control over the inputs its models learned from.

Three focused models became one production workflow.

With a shared reference in place, we could build and test the system in stages: check whether an image was usable, identify the cells, and rank the candidates for selection. Each stage had an acceptance threshold agreed before testing.

But meeting a test threshold wasn't a reason to hand over every selection decision at once. We used a shadow period in which the model ran alongside their scientists while they kept control of those decisions.

In under six weeks, we delivered the production system on their hardware. Because they now owned the replacement outright, they could eliminate the recurring vendor subscription. Their software subsequently became part of a licensed product running in pharmaceutical partners' labs.

We also recommended ongoing checks, including rescreening a sample of rejected candidates and periodic blind review sets. Their team budgeted for that oversight as part of operating the system.

Our acceptance tests needed a stronger design.

The system reached production, but a later review exposed weaknesses in how we'd measured the models. We had separated training and test images at the tile level, without keeping whole acquisition groups apart. Related images could therefore appear on both sides of the test.

We also hadn't measured expert agreement across the different image conditions, and one quality threshold could be met by a simplistic baseline. Clearing the agreed thresholds didn't answer every question we should have asked about performance.

So we brought those issues to the client, whose team reported addressing them internally. The gaps also changed our own approach: check independence, expert agreement, and a simple baseline before accepting the test design. The technical companion explains those lessons in more detail.

Their engineers kept building without an AI department.

The company owned the system, but its six receiving engineers had no AI specialists to turn to internally. So our handover included documentation and walkthroughs they could use to operate it and keep developing it.

Eighteen months later, they had extended the workflow to a new molecule class without hiring machine-learning or computer-vision specialists. The team had a foundation it could use beyond the original delivery.

We would make that handover stronger today by adding supervised operation and a period where their team runs the system while we observe. Their successful extension mattered, but it didn't remove the opportunity to make transfer more reliable.

How we built and measured it

Read the technical details for the method, evaluation, and handover.

We solved the definition problem that both earlier attempts inherited.

They came looking for better AI after two attempts had fallen short. But both attempts had inherited the same conflicting instructions, so we established a shared standard first. We rebuilt the data and models around it and delivered a system they could own and extend. The core AI reached production in under six weeks and later became part of a licensed product running in pharmaceutical partners' labs. The company had replaced its recurring subscription with software it owned, and its engineers went on to extend that software to a new molecule class.