How LiR Labs found microscopy experts for a niche no crowd can label

LiR Labs automates biological water treatment. Its predictive platform combines hardware and software into a digital operator: inline microscopes watch the microorganisms in a treatment plant continuously, and AI turns what they see into real-time decisions. Beyond wastewater, the same approach applies to bioreactors in biological manufacturing and to live cultures in food and beverage production.

Around 80% of the world's wastewater is treated biologically, by microorganisms in activated sludge. Their health decides whether a plant runs well, yet operators cannot see it change in real time. Samples go to the lab every now and then, so problems surface only once performance has already dropped.

• Treatment anomalies creep up unseen and cost money by the time anyone notices.
• Effluent quality drifts between lab samples; one exceedance of the discharge limits can mean fines and reports.
• Blowers run on fixed setpoints and overshoot the oxygen the biology needs, and aeration is often more than half of a plant's energy bill.
• A quarter of wastewater operators will retire in the next decade, and their experience leaves with them.

LiR Labs puts a microscope into the process and lets AI read it around the clock. The platform covers the whole loop:

• Understand: continuous, AI-driven microscopic analysis of the sludge instead of occasional lab samples.
• Act: software that helps run the plant cheaper and more reliably, based on what the biology actually needs.
• Anticipate: alerts that flag warning signs up to 24 hours before operations are disrupted.

The core of it is computer vision on microscope images. Those models are only as good as the people who annotate their training data: someone has to recognize, outline and classify the structures in every image correctly before a model can learn to do it on its own.

The work called for annotators who can read microscope images of activated sludge, the biology at the heart of wastewater treatment. That is not a skill you find in a general labeling workforce, and not one you can teach in a one-page guideline. LiR Labs defined what it needed precisely:

• Practical experience with activated sludge microscopy.
• Recognizing filamentous bacteria by their morphotype.
• Estimating the filament index of a sample.
• Telling bulking sludge, compact flocs and foam apart.
• Experience with image segmentation, because the structures have to be outlined by hand.

In-house, LiR Labs had a small team that could label, but not the capacity to scale. Hiring specialists one by one would have taken months for a skill set that only a handful of people in any one region have.

Pathwize turned LiR Labs' skills overview into a targeted call for experts across its network of students, graduates and specialists, reaching people by field of study instead of posting a generic job ad.

• Call for experts: one targeted call, shown to members whose field of study fits, from water science and microbiology to environmental and process engineering.
• First review: 100 applications came in, 84 of them complete and relevant enough to assess.
• Scoring against the criteria: every profile was checked against LiR Labs' hard criteria (activated sludge microscopy, filament morphotypes, filament index, bulking versus foam) and its preferences such as segmentation experience.
• Tiering: the strongest matches went into Tier 1, solid profiles that need a short onboarding into Tier 2, together with their weekly availability.
• Anonymized shortlist: LiR Labs received anonymized profiles to review with its ML team before meeting anyone.
• Practical screening: before work starts, each candidate labels a few real images (bulking sludge, a compact floc, foam), so self-assessment is checked against actual skill.

Out of 84 usable applications, 22 experts matched LiR Labs' criteria:

LiR Labs reviewed the anonymized shortlist with its team and tech lead. In a niche where most labeling vendors cannot staff a single qualified annotator, the pool for the pilot stood within weeks, and it met LiR Labs' criteria better than the team had expected.

The annotation itself starts once LiR Labs' images and labeling guide are ready for external work. The plan is the approach most teams with hand-segmented data end up with: label a base set in-house, train a small model that proposes masks, then scale the annotation with external experts who correct and refine them. The expert pool stays in place for that start, and it keeps growing.

• Write the criteria down first. LiR Labs' precise skills overview is what made a 100-application call sortable in days.
• Match by field of study, then verify. Degrees narrow the pool; a practical test on real images shows who can actually do the work.
• Tier the shortlist. A ready-to-start Tier 1 plus an onboardable Tier 2 gives a team room to scale without starting the search again.
• Get the data ready before scaling. Pre-segmentation by a small in-house model makes external expert time far more effective.