Machine Learning Engineer (Computer Vision) to Enhance ResNet Model + Implement Human-in-the-Loop (HITL)
Budget: $25 – $50 USD
I am looking for an experienced machine learning engineer with a strong background in computer vision to help us improve and scale an existing image classification system.
Our current model is based on a ResNet architecture and is capable of identifying bacterial colonies from images with fast inference (~1 second per image). However, classification still requires significant manual input, and we want to take the system to the next level.
Project Goals:
Improve the accuracy and robustness of our current ResNet-based model
Implement a human-in-the-loop (HITL) active learning system
Reduce manual labeling workload by intelligently selecting uncertain samples
Enable continuous learning from human feedback
Key Features We Want to Build:
Confidence/uncertainty estimation (e.g., entropy, margin sampling)
Automatic flagging of low-confidence predictions
Pipeline for sending uncertain images for human labeling
Integration of newly labeled data into retraining workflow
Model calibration (to address overconfidence issues)
Performance monitoring and evaluation pipeline
Current Setup:
ResNet-based image classification model
Dataset of bacterial colony images
Manual classification workflow
What We Need From You:
Evaluate our current model and pipeline
Design and implement an active learning / HITL system
Recommend improvements to model architecture or training process if needed
Help us structure a scalable training + feedback loop
Required Skills:
Strong experience with PyTorch or TensorFlow
Proven experience in computer vision (CNNs, image classification)
Experience with active learning / human-in-the-loop systems
Understanding of model calibration and uncertainty estimation
Ability to design production-ready ML pipelines
Nice to Have:
Experience with medical or biological image data
Familiarity with annotation tools (e.g., Label Studio or similar)
MLOps / deployment experience
Please share:
Examples of similar projects (especially active learning or HITL systems)
Your approach to implementing uncertainty-based sampling
Suggested improvements you would explore for our use case
We are looking for someone who can think critically about the system and help us significantly reduce manual effort while improving model performance.
Our current model is based on a ResNet architecture and is capable of identifying bacterial colonies from images with fast inference (~1 second per image). However, classification still requires significant manual input, and we want to take the system to the next level.
Project Goals:
Improve the accuracy and robustness of our current ResNet-based model
Implement a human-in-the-loop (HITL) active learning system
Reduce manual labeling workload by intelligently selecting uncertain samples
Enable continuous learning from human feedback
Key Features We Want to Build:
Confidence/uncertainty estimation (e.g., entropy, margin sampling)
Automatic flagging of low-confidence predictions
Pipeline for sending uncertain images for human labeling
Integration of newly labeled data into retraining workflow
Model calibration (to address overconfidence issues)
Performance monitoring and evaluation pipeline
Current Setup:
ResNet-based image classification model
Dataset of bacterial colony images
Manual classification workflow
What We Need From You:
Evaluate our current model and pipeline
Design and implement an active learning / HITL system
Recommend improvements to model architecture or training process if needed
Help us structure a scalable training + feedback loop
Required Skills:
Strong experience with PyTorch or TensorFlow
Proven experience in computer vision (CNNs, image classification)
Experience with active learning / human-in-the-loop systems
Understanding of model calibration and uncertainty estimation
Ability to design production-ready ML pipelines
Nice to Have:
Experience with medical or biological image data
Familiarity with annotation tools (e.g., Label Studio or similar)
MLOps / deployment experience
Please share:
Examples of similar projects (especially active learning or HITL systems)
Your approach to implementing uncertainty-based sampling
Suggested improvements you would explore for our use case
We are looking for someone who can think critically about the system and help us significantly reduce manual effort while improving model performance.