Bioavlee: 98% accurate software for automated bacterial colony counting

Bioavlee: innovators in automated bacterial colony counting

Bioavlee is a biotechnology company developing advanced hardware and software solutions for automated bacterial colony counting. Its proprietary laser-based imaging technology enables high-precision acquisition of bacterial samples, which are then analyzed using machine learning models to identify and count colonies with laboratory-grade accuracy. The solution significantly reduces manual workload while maintaining the level of accuracy required in highly regulated laboratory environments. Key insights:

98%

validated accuracy in image-based bacterial colony counting

50,000+

labeled images in the training dataset

4 weeks

to develop and validate the first production-ready MVP

The challenge: scaling accuracy in a regulated environment

Bioavlee’s core expertise lies in laboratory instrumentation and optical hardware. While the initial software performed reliably in controlled, small‑scale laboratory settings, scaling the system exposed significant challenges. As Bioavlee operates in a highly regulated environment, any degradation in accuracy or reproducibility posed unacceptable risk. Early attempts to scale revealed several issues common to real‑world medical datasets:

label noise introduced during manual annotation;

class imbalance between bacterial strains;

data drift as new samples were introduced.

At scale, model performance dropped to ~60% accuracy on datasets exceeding 50,000 images - well below acceptable thresholds for production use. Bioavlee needed a partner capable of building reproducible, scalable, and validation‑ready ML software, with transparent evaluation processes suitable for external audits and regulatory review.

The solution: cross‑functional ML and domain collaboration

Bioavlee partnered with CODEWAVE to redesign and harden their image recognition software. The collaboration began with in‑depth Exploratory Data Analysis (EDA) workshops involving both CODEWAVE’s software engineers and Bioavlee’s domain experts. This phase established a shared understanding of laboratory constraints, data characteristics, and regulatory requirements. CODEWAVE then executed a structured improvement program combining:

disciplined data curation and relabeling workflows;

close collaboration with domain experts;

deep neural network (DNN) optimization;

statistically rigorous validation and hold‑out testing.

Within the first four weeks, the joint team increased model accuracy from 60% to 82% and delivered a roadmap toward full production readiness.

"CODEWAVE quickly became an integral part of our organization. Their team combined deep ML expertise with the ability to truly understand our laboratory domain. The value of applying neural networks became clear immediately."

Michal Wronecki/

CEO, Bioavlee

Used technologies:

Deep Neural Networks (DNNs)

for image‑based colony detection and counting

On‑premise NVIDIA GPU servers

for sensitive and validation‑critical workloads

Amazon EC2 (p2.8xlarge, Spot Instances)

for scalable and cost‑efficient training

Machine learning

with strict separation of training, validation, and test data

The results: repeatable ML validation and 98% accuracy

Over the following six months, CODEWAVE and Bioavlee collaborated on continuous model refinement, data quality improvements, and evaluation methodology. Key outcomes included:

a repeatable training and validation pipeline;

strict hold‑out testing to control overfitting;

transparent, reproducible metrics suitable for external review.

This iterative process increased model accuracy from 60% to 98%, while ensuring statistical robustness and long‑term maintainability. Additionally, the introduction of hybrid training workflows, connecting on‑premise GPUs with AWS resources, enabled scalable experimentation without compromising reproducibility or cost control.