Sometime back in 2011 or 2012, we built a Machine Learning system for camera-to-camera handoff of people detection. So, our ML system used ceiling-mounted security cameras (often fisheye) in retail stores and malls to identify and track people. A handoff algorithm was necessary when people walked out of one camera and into another. It worked perfectly well. This experience was quite insightful, and I learned that you can do anything you put your mind to, even long before the age of Transformers.
We wanted to analyze the top ranking pages on Google for given search terms and keywords. We then wanted to reverse engineer these to see how Google decides what to rank. What structure the pages were, what semantically related terms were used, how optimized for SEO those pages were etc. It was initially hard filtering out the "noise" semantic terms in each of the pages and reducing the list down to just those that really matter. We had to work with various NLP tools and fine tune the models in order to create software that could reliably work across niches. We learned that having a human in the loop to for testing and validating results was vital.
I worked on a blockchain project that tokenized carbon credits, where AI was used to verify the legitimacy of climate offset projects through satellite imagery and data from IoT sensors. The biggest challenge was model reliability in remote and diverse geographies. We used image segmentation and anomaly detection models to spot inconsistencies in reported versus observed data. This project taught me how AI can serve as a truth layer in blockchain systems, especially in sustainability and ESG applications.
One project I'm particularly proud of involved developing a predictive analytics model for a retail client aiming to optimise inventory management. The challenge was the vast amount of unstructured data from various sources, including sales history, customer behaviour, and seasonal trends. To tackle this, I implemented a combination of natural language processing and time series forecasting. The initial model struggled with accuracy due to data noise and missing values, but through rigorous data cleaning and feature engineering, I improved its performance significantly. This experience taught me the importance of data quality and the iterative nature of machine learning. Collaborating closely with stakeholders also highlighted the value of aligning technical solutions with business goals. Ultimately, the model reduced excess inventory by 30%, showcasing the tangible impact of machine learning in driving business efficiency.
One machine learning project I'm particularly proud of was developing a predictive model for customer churn for a SaaS company I worked with. The goal was to use historical customer behavior data to predict which users were likely to cancel their subscriptions, so the team could intervene proactively. Challenges faced: Data quality: The data we had was messy—missing values, inconsistent formats, and a mix of structured and unstructured data. Cleaning it up took more time than expected, but it taught me the importance of data preprocessing as a foundational step. Feature selection: Choosing the right features was tricky. We had an overwhelming amount of variables (user activity, customer support interactions, payment history, etc.), and we had to figure out which ones were truly predictive of churn. We used feature importance techniques and cross-validation to narrow it down, which was a learning experience in optimizing models. Model tuning: Initially, our model's accuracy wasn't great. We tried several algorithms—Logistic Regression, Random Forests, and XGBoost—before landing on a hybrid approach that balanced precision and recall. Tuning hyperparameters was key to improving the model's performance. What I learned: Data is king: No matter how sophisticated the algorithm, if the data isn't clean and well-structured, it won't work. We spent as much time on data wrangling as we did building the model. Model interpretability: It's not just about finding the best model but understanding why it works—for both technical and non-technical stakeholders. We focused on creating a model that could be easily explained and acted upon, which was essential for business buy-in. Iterative improvements: Machine learning models are rarely perfect from the start. It's about continuous iteration—fine-tuning, monitoring, and tweaking over time to improve results. The project ended up being a success, with the model helping reduce churn by 15% within 6 months, and the company was able to save thousands in lost revenue. It was a reminder that, in machine learning, the process is just as important as the final result.
A machine learning project I am particularly proud of is Bio-LUSH, where we tackle the challenge of sustainably utilizing underexploited biomass feedstocks such as hemp hurd, nettle, forest residues, and seagrasses for high-quality fiber extraction. One of the key challenges we anticipate is the extensive characterization and data collection required for these diverse biomass feedstocks to ensure robust machine learning outcomes. We plan to carefully curate comprehensive datasets from advanced characterization of feedstocks, fibers, and bio-based products to enable accurate predictions and optimal design of sustainable materials. Through Bio-LUSH, we will demonstrate how machine learning tools can effectively predict fiber properties and optimize processing methods, facilitating the creation of innovative bio-based materials such as edible packaging, antibacterial textiles, and 3D printable bio-composite filaments for automotive interiors. We expect to gain significant insights into the immense potential machine learning holds, not only within traditional tech sectors but also significantly impacting the bioeconomy by enhancing sustainability, circularity, and efficiency in material development. Bio-LUSH illustrates that machine learning is essential for advancing circular and sustainable practices in biomass utilization.
I once worked on a machine learning project where we developed a recommendation system for an e-commerce platform to personalize the shopping experience. We started by collecting and preprocessing data, including user browsing history and purchase patterns. Using Python and TensorFlow, we built a collaborative filtering model to predict user preferences. After training the model, we integrated it into the platform's backend using REST APIs. The system was able to make accurate recommendations, leading to increased user engagement and sales. Throughout the project, we collaborated with data scientists and engineers to ensure seamless integration and performance. This experience not only enhanced my technical skills but also taught me the importance of teamwork and clear communication in delivering successful machine learning solutions.