Aron T.

Given my previous experience with voice bots, I was invited to this project as an expert consultant. I gave insight on prompting, voice2voice networks, transcription handling, and real-time information injection. The project uses ElevenLabs (which sadly isn't in the technologies list). My role here is not a constant load, more like being involved occasionally when there are uncertainties around technology or user experience. My role was more crucial in the beginning of the project when obviously there were more decisions to make. As I'm writing this, the project is about to go live, so my work here is mostly done.

Adnexo is operating a system that tracks trucks, and generates road usage reports to the Swiss government. Due to legislatory changes, their usage more than quadrupled in a short period, and their system was unable to handle the load. My first task was to stabilize the system, and make it performant again, by optimizing both source code and infrastructure. But the load was still growing, so the CTO decided to move over to AWS, which meant moving 7 billion records within a very small timeframe, since the system is real-time, so it can't be down for long. I also did a major refactor of the core of the system, making it parallel and scalable, and many new features and upgrades.

• Served in a consultancy and architectural advisory role for a Norwegian startup focused on large-scale document intelligence.
• Led the design of a scalable and robust architecture capable of handling complex operations on data indexed from sources such as Google Drive and SharePoint.
• Provided strategic guidance on integrating and orchestrating LLM-based components, ensuring the system could perform composite operations beyond the capacity of any single model.
• Advised on the backend implementation built in Node.js and a React-based frontend, aligning system performance with business-critical requirements.
• Developed strategies for multi-step reasoning, context retention, and intelligent aggregation of results across heterogeneous document sources.
• Ensured architectural decisions supported extensibility, enterprise readiness, and seamless user experience under high load and varied use cases.

• Architected and developed the backend of an innovative voice-driven chatbot platform designed to simulate difficult real-life conversations for training purposes.
• Built the backend system entirely from scratch in pure Python with Django, drawing only loosely from legacy concepts while implementing a fully modern, maintainable codebase.
• Integrated LiveKit to enable real-time voice calls, forming the foundation for interactive voice simulations.
• Developed support for complex conversational scenarios, including sales resistance handling (e.g., selling gym memberships to hesitant prospects) and delivering constructive feedback to underperforming employees.
• Implemented robust session management, dialogue state handling, and voice processing workflows to support realistic, adaptive conversation flows.
• Participated in product planning and prototyping to align technical capabilities with the behavioral training goals of the platform.

• Designed and developed an advanced Retrieval-Augmented Generation (RAG) based chatbot that encapsulated and delivered the expertise of Alistair Brownlee, two-time Olympic gold medalist triathlete.
• Engineered the system for hallucination-free, highly accurate responses, emphasizing traceability and explainability of generated content.
• Built the solution with dynamic content update capabilities, allowing the chatbot’s knowledge base to evolve with new information over time.
• Implemented the backend architecture to support real-time queries and seamless integration with the Brownlee Fitness platform.
• Ensured robust performance and scalability while maintaining user-friendly, conversational interactions tailored to the needs of fitness enthusiasts.
• Delivered a production-ready, long-lasting solution that remained in active use on brownleefitness.com.

• Conducted a rapid, high-stakes codebase audit following the abrupt departure of a developer under conflict, ensuring no malicious changes or damage had been introduced.
• Performed a thorough review of a large-scale Django (Python) and TypeScript codebase, including both backend and frontend components.
• Verified code integrity, evaluated recent commits, and identified potential vulnerabilities or suspicious modifications.
• Delivered a comprehensive assessment to the client, confirming the security and stability of their technical assets.
• Advised on best practices for codebase security, version control hygiene, and developer access policies to prevent future risks.

• Solely developed a Python-based desktop application for advanced image and sunlight analysis;

• Collaborated closely with a Product Owner to define project scope, functional requirements, and technical feasibility;

• Engineered algorithms to calculate the precise moment of sunrise and sunset at specific geographic locations, factoring in complex variables such as terrain, buildings, and vegetation;

• Applied mathematical modeling and astronomical computation using libraries like NumPy, SciPy, and specialized astronomical toolkits;

• Designed and implemented a robust backend architecture capable of handling high-resolution spatial and visual data;

• Conducted rigorous testing and validation of image analysis algorithms to ensure accuracy in varying environmental conditions;

• Optimized application performance and accuracy by refining computational geometry and time-based prediction logic;

• Ensured a user-friendly desktop interface, maintaining a balance between technical functionality and usability.

• Acted as a supervisory and managerial lead overseeing the performance and collaboration of an outsourced Polish software development team;

• Identified and addressed critical issues in team cooperation, communication breakdowns, and inefficiencies in project execution;

• Advised the Product Owner on assessing productivity concerns and pinpointing quality bottlenecks in the development process;

• Redesigned the team’s workflow by introducing agile ceremonies and structured communication channels to facilitate transparent and effective data exchange;

• Conducted comprehensive code reviews to improve software quality and maintainability;

• Mentored developers on best practices and instilled a culture of accountability and continuous improvement;

• Oversaw the successful transformation of team dynamics and delivery standards, leading to significant improvements in both development speed and code quality.

I worked for OpenAI as an AI trainer, my role was mostly focused on the Chat Completion feature, introduced in ChatGPT 4. I trained ChatGPT for using various programming languages, mostly Python, to detect when the use of programming is required to answer the question, and how to verify if the solution given by the program is correct. I was part of a group of about 4000 people.

• BMAT is processing music usage all over the world, almost a petabyte of data every day
• The work is about the monetization of music usage by DSPs
• Working with the data of Byte Dance, YouTube, SoundCloud, and a lot more
• The product is based on Django, MongoDB, multiple queuing components, data warehousing, and AWS, with Vue.js on the front-end
• Enjoy the small company size and the flat structure, mainly because I get to meet the customers every week
• My out-of-the-box ideas and my Scrum expertise are both welcome in the company

• Founded a startup for socially assistive robots speaking Hungarian
• Lead a team of 5, wrote code from the ground up, using ROS, Python 3, and Flask
• Integrated face detection and face recognition using OpenCV, pose detection, and emotion detection
• Trained Deep Speech model for Hungarian language recognition, an NLU for Hungarian grammar
• Had various kinds of robots, including a virtual human 3D model created in Unity with mocap animations
• Unfortunately, scaling up was not possible in Hungarian

• Facilitated Scrum ceremonies (sprint planning, daily stand-ups, backlog refinement, sprint reviews, and retrospectives) as a Developer/Scrum Master before transitioning to a pure development role.
• Designed and implemented the data-collection and processing component for an indoor UWB-based localization system, deploying Python 2 code on Raspberry Pi devices.
• Architected and developed a web-based user interface from scratch using Python 3 and the Bottle framework.
• Containerized application components with Docker to streamline deployment and ensure consistent development environments.
• Integrated PostgreSQL as the back-end database for storing and querying localization data.