Tomislav Peharda

I worked on a project focused on developing a similarity search tool to handle millions of data points. The task involved creating an innovative solution for users to query data and find similarities using anchor assets or textual queries.

The challenge was creating a tool to manage vast data while providing accurate search results efficiently.

Leveraging my skills in Node.js, Machine Learning, GraphQL, and Elasticsearch, I contributed to the design of this semantic search tool. I utilized lightweight technologies, ML algorithms, and search optimization algorithms to ensure the tool could handle the data load and provide accurate results.

This tool has significantly improved data querying speed and accuracy, providing users a more streamlined and effective way to find similar data points.

My team and I worked on a new microservice designed to support the processing and storing of billions of structured and unstructured assets. This involved the creation of an ETL pipeline, API, and several SQL and NoSQL storages to enhance data querying and search capabilities.

Our primary objective was to design a system capable of handling high-volume data efficiently. We aimed to minimize processing time while maximizing search speed and flexibility in data querying.

The impact of our efforts was significant, resulting in a scalable, performant service capable of efficiently processing and storing billions of assets. This increased data availability and introduced new ways of querying data, significantly improving search capabilities.

A project to migrate their existing Python microservices to Node.js architecture. The key focus was to develop a Nest.js-based framework that abstracts the complexities of the NestJs framework, tailored to align with the client's requirements.

The challenge was in the migration and creation of a simplified yet robust Node.js architecture from existing Python microservices without disrupting the client's operations. The company had a framework for bootstrapping Python applications and wanted a framework similar to Node.js.

I led the development of a Nest.js-based framework that simplified the complexities of the Nest.js framework. The client has successfully migrated a couple of their Python microservices to Node.js using the developed framework.

As a machine learning engineer, I utilized AI to optimize the efficiency of legal case analyses by identifying similar legal matters.

The project aimed to simplify the complex and time-consuming process of case comparison to better predict the potential outcomes of ongoing legal matters.

In this project, my team and I developed innovative weighting logic for each case component. I utilized AI, Python, Machine Learning, and TensorFlow skills to employ cosine similarity techniques for case identification.
Introducing this AI-based system significantly improved the efficiency and accuracy of legal case analysis. This not only streamlined the law firm's operations but also enhanced the decision-making process for attorneys, leading to a profound positive impact on the firm's overall workflow.

This project involved the development of plugins and modifications to extend the gameplay of popular video games like CSGO, WoW, PUBG, and more. This was accomplished by reading game process memory and providing additional information to the user.

As a full-stack developer, I leveraged my skills in reverse engineering, data visualization, Lua, Redux, Docker, React, Node.js, and JavaScript to develop custom plugins and modifications that improved the player game experience.

This project significantly enhanced the overall gaming experience, leading to increased user engagement and satisfaction. The additional information provided by the plugins and modifications allowed users to get more out of their gaming experience.

In the venture capital industry, the team at Growth Warrior Capital and I developed an AI-powered pitch deck builder. The core objective was to elevate and improve user responses to align better with industry standards using AI.

The main challenge was that user responses often lacked the necessary quality and adherence to industry standards, which led to ineffective pitch decks.

As a deep learning engineer, I utilized my Python, AI, machine learning, NLP, GPT/OpenAI, and data modeling skills. I contributed to developing AI features that validated user inputs, provided constructive feedback, generated improved examples, and educated users on why these examples were superior.

The result was a significant enhancement in the quality of user responses, leading to more effective and industry-compliant pitch decks, which, in turn, helped users secure more funding opportunities.

With my team, I worked on a dashboard providing insights into metrics of autonomous vehicle rides, spotlighting areas for improvement.

The key challenge was to create a comprehensive system that would effectively analyze and present ride data to identify improvement areas.

This dashboard not only eased the data interpretation process but also provided valuable feedback, leading to significant improvements in the autonomous vehicles' performance.

I led the design and development of a robust tool to analyze ML models. This tool was designed to delve into datasets and enhance ML model accuracy by providing insights into what cases the model might not be capable of predicting so well.

I developed a solution by using my skills in Artificial Intelligence (AI), Node.js, React, GraphQL, Systems Architecture, and PostgreSQL. I led the project, conceptualizing and executing the development of a tool to analyze ML models by assessing datasets.

The tool helped numerous enterprise consumers analyze their ML models, enhancing performance more efficiently. This included the application of different software engineering and ML engineering algorithms (such as TSNe, similarity search, etc.) to gain insights into the dataset.

This project involved developing a predictive model to estimate the dollar value of a legal matter. As a deep learning engineer, my role was pivotal in leveraging data science and engineering techniques to provide attorneys with a tool to assess potential case value.

The main challenge was to quantify the potential value of a legal matter, a task traditionally dependent on subjective human judgment.

I used my Python, AI, machine learning, NLP, data pipelines, and data analysis skills to create and test various traditional machine learning and deep learning algorithms. The goal was to design a system capable of making accurate predictions based on historical and contextual data.

The result of this ongoing project is a sophisticated predictive tool that reduces uncertainty and enables more informed decision-making in the realm of legal affairs.

As a data scientist, I worked on the application for providing better financial security for lower-income US citizens. The recommendations were given based on the various characteristics of an individual. As part of my role, I worked on the budgeting features such as scoring an individual's current financial situation, forecasting expenses, analysis of transactions, and more.

I worked on a social network for investors to enable them to share potential opportunities where they might collaborate. The network was Twitter-like, where users could share updates with various types of content. Users were provided with tools to personalize their profiles and provide insight into their background and the opportunities of their interest. The work also included a live messaging system.

A video content marketing company engaged me to employ my skills as a data scientist in improving the quality of their video content review process. I achieved this by analyzing their extensive video media database using various Machine Learning and statistical methods.

The client had a vast amount of video media data, along with various metadata that had been reviewed over time. The challenge was figuring out how to use this data to enhance the quality of to-be-reviewed video content.

As a data scientist, I employed various skills to analyze the dataset. This analysis aimed to help the company better target what video content to focus on.

By leveraging AI and machine learning, the client could now prioritize their video content based on the insights derived from the data analysis, thus leading to an improvement in their overall video content quality.

I worked on the payment features of online software for dental practices. The dental practices would use available desktop software to track patients and treatments. However, these applications had no payment integration or web interface. As part of the project, I worked on functionalities for the web application connected to the desktop application that enabled processing the patient treatments end-to-end.

I worked on a pipeline based on manually labeled and reviewed YouTube videos and would attempt to find similar videos and channels through the construction of queries that targeted titles and keywords. A user would select various criteria to generate the list of already reviewed videos and use data from these videos to generate search queries.

Going further, YouTube would return relevant videos and channels based on the search query. However, the results might contain outliers; thus, the second part of the pipeline was to verify the quality of the returned analysis, which included the analysis of metadata through NLP techniques.

Acted as a member of the Faculty of Organization of the Varaždin artificial intelligence laboratory. As part of the laboratory work, I collaborated on writing scientific papers on AI, wrote coding snippets, and was engaged in a project that focused on developing an orchestration platform to support multi-agent systems.

I worked on a tool that enables building simple internal applications, mainly targeting analytic tools. Users would be able to construct the UI the way they want within the boundaries of the components provided. Users could connect their applications to data sources such as Snowflake, MySQL, etc.

I worked on a project that provides insight to an adequate audience for medicine trials based on the survey data collected from millions of people. The surveys contain questions about an individual's health, which were then used for querying based on various criteria. The criteria could be anything from demographic and geographic to health conditions. The pharmacies would then get insight into what percentage of people are adequate for trials and may invite them for the experiments.

For my master's thesis, I have developed an intrusion detection system—a neural network (model) that inspects network traffic requests in order to determine whether a request is an intrusion or a valid request (the output of the model is a class of an intrusion).

The model is developed in TensorFlow, and it utilizes a Random Forest concept to minimize the number of attributes needed for the computation. It is based on the KDD1999 dataset. Different methods, utilities, and metrics have been used to examine the dataset and provide the most accurate solution.

In my thesis paper, I've used various data science approaches to examine the quality and diversity of the KDD1999 dataset. I also have shown how different neural network configurations (number of layers, number of neurons, activation function, regularization, and so on) impact output precision.

Technologies: Pandas, NumPy, Python, TensorFlow (Keras API), Random Forest

I've worked on various projects for a range of clients including a law firm, a logistics company, and a gym.

Law Firm
The purpose of this web application (which also worked on a phone) was to enable the law firm to schedule and track appointments. Each appointment would have a detailed history of events. A user was also able to set notifications for each appointment.

Logistics Company
This web application was used to track drivers, trucks, trailers, destinations, and the like, however, the primary focus of the app was to have a weekly preview of an each user's schedule on a dashboard that is also highly interactive thus simplified the process of creating weekly schedules at the organizational level.

Gym
For this client, I built a web app that was meant to be used to track each user's subscriptions. A gym would have an RFID reader at the entrance requiring each user to scan their subscription card. That way, the owner would, in real-time, have access to the number of users at the gym — seeing whether any of them has not yet paid their subscription or such. The application also implemented various analytics previews.

Technologies: PHP, HTML, CSS, JavaScript, MySQL, jQuery, Bootstrap

I've also reverse-engineered games with the Cheat engine, x96dbg, IDA disassembler, and developed both internal and external modifications in C# and C++. A strong knowledge base of x86 architecture and Assembly was the main priority.

Some of the main activities in the process are: analyzing and finding proper memory addresses, examining the Assembly code and making changes that would affect the game behavior, developing modifications that override memory locations and also invoke game functions in order to achieve the desired functionality, bypassing basic security mechanisms (some games would detect if the memory location value gets changed externally, other security mechanisms include updating multiple memory locations with the same value).

I developed an open-source modification in C/C++ for CoD2 to improve their competitive gameplay which was officially used in every game league that supported CoD2 (it is estimated that around 20,000 players used the modification). With the introduction of the modification, the game activity on a monthly basis was largely increased.

ServerBot was a functionality (or even a protocol) within the modification to enable a web application to listen to the server events in real-time and then preview them. The web application that I've built to make use of ServerBot was able to listen to 5 game servers simultaneously with ability to preview the events in real-time or with a delay.

During my mobile development journey, I created a cross-platform application for storing user credentials on a phone and use the credentials to automatically sign-in into web applications. The primary focus was put on application security and data encryption.

I've worked on a multiplayer (peer-to-peer) game that utilized match-making system support through a central server to enable players to find an opponent automatically (also, the central server was used to store a user's stats). Once an opponent has been found, the communication between users no longer went through the central server but sockets (peer-to-peer connection). TCP hole-punching was used for players to establish a connection.

Technologies: Java, Corona SDK, Lua, Hole Punching, MySQL, SQLite

For a JavaScript-based market, I developed highly customizable JavaScript components out of which also included WordPress support. The most successful app I've developed was a scheduler that read events from a JSON file. The component provided multiple language support and on a date click, it would preview all events for that date with location (Google Maps integration) and time.

The other two components I've worked on were a 3D item slider and a carousel. The focus of the 3D slider was on shopping items—the items would be rotating (in different manners) and on a click, information for the selected item would be shown (a user could define information themselves). The carousel had similar use-case as the 3D item slider, however, the entity of a focus was not shopping items but rather images (therefore, usage, animations provided and implementation were fairly different from the 3D item slider).

Technologies: JavaScript, jQuery, HTML, CSS, WordPress, JSON