Machine Learning for Bioavlee
About the Client
Bioavlee is a biomedical startup building automated bacteria recognition for labs and industry. Working in a regulated setting, the team needed a production‑grade classifier that keeps high accuracy as data grows and varies. They also required reproducible, auditable training with clear data/model lineage for external validation while fitting existing lab workflows and keeping training costs predictable.
Key Insights
Accuracy restored
From 60% back to >98% on a diverse 50k+ image dataset.
Data‑first sprint
4 weeks of EDA and workshops produced an 82% prototype and a clear improvement roadmap.
Hybrid training
On‑prem servers plus AWS bursts for cost‑efficient scaling.
Audit‑ready
Formal validation pipeline aligned to external review requirements.
The Challenge
The previous solution reached about ninety percent accuracy on a small dataset but dropped to about sixty percent as the dataset grew beyond fifty thousand images. The decline revealed class imbalance, label noise, and domain drift, putting delivery timelines and third party validation at risk. Bioavlee set a clear goal: reach a recognition rate of ninety eight percent on the full dataset while keeping training scalable, reproducible, and cost efficient.
The Solution
CodeWave ran a structured program that combined data analysis, domain collaboration, and disciplined MLOps to restore accuracy and scale the system.
Discovery and prototypingA four‑week EDA and joint workshops surfaced label noise, class imbalance, and data drift; a deep neural network prototype reached 82% accuracy and produced a prioritized improvement plan.- Iterative validationOver six months we worked with Bioavlee’s experts on data curation and evaluation, refined the model architecture, and introduced a repeatable training pipeline with strict hold‑out testing to control overfitting and ensure reproducibility.
- Hybrid training at scaleTraining ran on a hybrid setup on‑premise GPUs for routine cycles and AWS for elastic bu
Technology
Deep Neural Networks
On-Premise Servers with Nvidia GPUs
Amazon EC2 Instances
(p2.8xlarge, Spot Instances)
Exploratory Data Analysis (EDA)
They immediately became part of our organisation and provided truly cross-functional specialists, who were able to easily ingest the adequate level of domain knowledge to communicate with the rest of the team. They explained the theory clearly and in simple words and we were able to immediately see the benefits of applying neural networks to our problem.
Key Results
98%+ accuracy on the full dataset
reverses the earlier drop from about 60% on 50k+ images.
Stable at scale
new classes and devices added with no quality degradation.
Reproducible and audit‑ready
fixed hold‑out testing, versioned data/models, clear lineage.
Cost‑efficient training
hybrid on‑prem + AWS bursts for predictable spend and fast iteration.
The Summary
CodeWave restored Bioavlee’s recognition performance from about sixty percent to more than ninety‑eight percent and delivered a reproducible, audit‑ready pipeline. The solution scales with new data, keeps training costs predictable through a hybrid setup, and is ready for external validation.
Let’s create a success story together.
contact
Poland: CodeWave sp. z o.o. Opolska 13, 52-010 Wrocław, Poland
+48 539 019 430, hello@codewave.eu
US: CodeWave LLC 16192 Coastal Hwy, Lewes, DE 19958, United States
+1 (631) 909 5771, hello@codewave.eu
Codewave. All Right Reserved 2025
