Marcin Bogdanski

A game-prototyping startup needed high-quality captions for over 1 million mobile game screenshots. Existing captions, scraped from app stores, were largely unusable: only about 10% bore any relation to the actual image content.

I defined a 6-category evaluation rubric (layout, characters, UI elements, camera angle, lighting, and artistic style), manually labeled 50-100 reference images, and benchmarked six vision-language models (Florence-2, LLaVA, MiniCPM, Phi-3-vision, and CogVLM2). I containerized the selected model, integrated it into the existing Docker and RunPod infrastructure, and processed the full dataset on a 4×4090 node. The pipeline was then deployed to production for ongoing ingestion.

For the generation, I built a PixArt Sigma fine-tuning pipeline using HuggingFace Transformers, implementing custom dynamic batching by image size from scratch. I conducted validation with multi-day test runs and handed off for production training.

The task was to analyze a large database of call center transcripts (over 1 million) to improve transcript quality and evaluate applied sales strategies (e.g., up-selling and cross-selling) for training purposes.

I curated a training dataset, evaluated multiple model architectures (BERT, encoder–decoder transformers, and others), selected the top-performing approach, and implemented training on an eight-GPU AWS node, running multiple experiments simultaneously. I packaged the trained model in Docker and deployed it to a Nomad cluster for production, processing approximately 5,000–10,000 transcripts per day. Compared with earlier manual methods, the project reduced costs by approximately eighty percent in the first year.

A Cisco-partnered startup needed real-time store occupancy tracking for COVID-19 compliance. I proposed an alternative to the client's original in/out counting approach: count people currently visible across existing security camera feeds, with masking to avoid double-counting from overlapping views.

The main challenge was that there was no API access to camera feeds—only a web interface. I engineered a workaround using Docker containers with virtual framebuffers, capturing frames via the browser's screen grab feature. The pipeline integrated two detection models to handle mixed standard and fisheye cameras.

The system processed 50-100 cameras across six pilot stores with a latency of 2-4 seconds, achieving an accuracy of over 90% against manual ground truth.

TECHNOLOGIES
Python, Docker, OpenCV, AWS, object detection models, and browser automation.

PROJECT DESCRIPTION:
Parking signs in Australia are notoriously complicated. The project was to build a proof-of-concept vision system to detect and recognize signs with a mobile phone and let the user know whether they can park at that location.

TECHNOLOGIES:
The system uses a customized state-of-the-art multi-stage neural network for detection. The main challenge was working with a tiny data set and a large variety of backgrounds, lighting and weather conditions and obstructions. We deployed the system for production on Amazon EC2

This project won first place and a monetary prize of $20,000 at the Melbourne Knowledge Week. I was fully responsible for the AI technology and the demo was delivered by the customer.

PROJECT DESCRIPTION:
Stock traders need to process a large quantity of complex information in real-time to be able to compete. The project was to analyze input PDFs and present to the user a short 2-3 sentence summary.

TECHNOLOGIES:
During the research phase, we experimented with both extractive and abstractive text summarization, sentiment analysis, and translation techniques to get all the pieces necessary for the final product. Decoding raw PDFs was a significant challenge as well.

PROJECT DESCRIPTION:
Smart home and IoT technologies offer huge potential savings in heating, power, air conditioning, and predictive maintenance. Combining sensor data from thousands of properties, we could identify actionable insights for occupants and landlords.

TECHNOLOGIES:
Houses were equipped with a variety of sensors like power, temperature, humidity, and CO2. We used a range of models like ARIMA, Gaussian processes, neural networks to detect latent variables (e.g., occupancy status), predict behavioral patterns (e.g., heating requirements), and risk of issues (e.g., hidden mold).

PROJECT DESCRIPTION:
This project is a proof-of-concept neural architecture for automated fault detection on high-resolution images of wind turbine blades. I provided know-how via ongoing consulting as well as developed actual software for data pre-processing and model training and testing. The customer provided images and preliminary labels in a raw format. A large part of the project was to browse, clean, select, and label data provided so it can be feed into the machine learning component.

TECHNOLOGIES:
Due to the unique technical challenges that we encountered, we developed two non-standard components:
1. A completely custom preprocessing pipeline to handle the characteristics of the input data.
2. A far-reaching optimization of a neural network subsystem to allow for fast training time.

PROJECT DESCRIPTION:
In this project, I worked as the principal software developer responsible for designing and implementing a self-driving module for an unmanned ground vehicle (UGV). The robot needed to navigate with GPS for predetermined routes where an obstacle avoidance module would be responsible for detecting unexpected obstacles like pedestrians, parked vehicles, and similar challenges (the robot would use walking paths).

TECHNOLOGIES:
The primary sensor was a front-facing monocular RGB camera (i.e., a webcam). Images from the camera were processed by two independent neural networks.

1. YOLO for object detection and localization—pedestrians, parked vehicles, and more. The network was trained on both a pre-existing dataset as well as a mix-in of our own training data.
2. SegNet was used for detecting if the ground in front of the vehicle is drivable (e.g., pavement, tarmac). The network was trained on cityscapes with a few added images.

PROJECT DESCRIPTION:
A CNC router is a machine that uses a rotational tool bit (drill) to remove material from a solid block to manufacture a target part. The tool bits wear down and technicians often forget to check and replace them. This system automatically takes a picture of the tool bit before the job has started and feeds it into a convolutional neural network to evaluate the current wear of the tool bit.

TECHNOLOGIES:
Image classification is performed with DenseNet-201 which showed the best performance out of the attempted architectures. CNN was initially trained on ImageNet and then trained further. The classifier was trained on a target dataset with heavy data augmentation and applied regularization. Afterward, the training network was able to detect cracked, chipped, or overheated (changed color) tool bits.

The system is currently operating in a machine shop at an aerospace manufacturing facility.

PROJECT DESCRIPTION:
The purpose of this project was to assess in detail the performance of a swarm of 200+ mining and support robots. The swarm would start in containers, then deploy solar panels, build a surface base, excavate tunnels, process raw materials, and finally wind up in an operation. The complexity of the project was similar to a simple strategy video game.

TECHNOLOGIES:
The simulation is physics-based and built in a Unity3D game engine along with a set of plugins for dynamic volumetric terrain (so that the robots can freely excavate). An individual robot AI manages energy, navigation, task queues, and more. The swarm AI manages task allocation, robot coordination, excavation orders, and similar.

At any point, a user can override the high-level strategy or take over full control over an individual robot.