Projects, Notes, Decisions

In this project, the robot has two important safety protocols: 1- The robot can be teleoperated from the mobile application we developed. The mapping is also done by this mobile application. After the mapping is done, the danger zones that the parent wants the child to avoid are marked by the parent and these coordinates are sent to the robot. While running using nav2, in the normal use case, robot uses its LiDAR data and constantly looks for the distance between the child and the danger zone and if some conditions are met, the robot moves between the danger zone and the child. It keeps the distance from the child using nav2 and local costmap inflation. 2- Robot has weights of a fine-tuned YOLOv8n. Using this, robot can detect if the child is holding a knife or bleach. In both protocols, a push notification is sent to the parent and it is logged in the database. For the robot to be interactive, there is an AI-powered conversational API integrated into the system, prompted to interact with and engage the child. These conversations are also logged in the database so they can be monitored through the mobile application.

If wanted, robot can be teleoperated by the parent anytime. To deactivate the nav2 navigation, there is a finite state machine implementation behind the scenes, both in the active blocker protocol and manuel control.

Stack: Python, C++, ROS2 Humble, OpenCV, SLAM, Nav2, octomap, Gazebo

Currently in development.

A website that the users can upload their datasets, visualize the data, clean the data and choose a model that best fits their data and save the model.

Stack: Python, FastAPI, React, TailwindCSS, scikit-learn

This project is developed to help users refresh their knowledge every day. They can solve the daily questions and level up their characters/classes and as users progress to higher levels, the questions become more difficult. It is designed to be only client-side, no backend needed. There are only API calls to the LLM to generate questions and inspect the user's github profile.

Stack: React Native

Project will uploaded in the github soon.

This project contains 3 collaborator agents and one supervisor agent to make podcasts from the given inputs from the frontend. The user can set characters that are going to talk and set the topic to be talked about. This is turned into a prompt for the supervisor agent that lies in the agentcore. After getting the prompt through Amazon API Gateway, the agent is using its MCP servers to perform web-assisted character analysis. After setting the characters, the supervisor coordinates the agents (using A2A that has been implemented via Strands framework) to make a script and then that script is getting turned into an MP3 file via a MCP server that uses Amazon Polly.

Stack: agentcore cli, Python, ddgs, Strands, bedrock-agentcore, AWS services (S3, DynamoDB)

Fine-tuned a Turkish language model using Parameter-Efficient Fine-Tuning (LoRA), focusing on improving its ability to explain mathematical concepts clearly in Turkish. The model was evaluated primarily through a human-in-the-loop approach, where generated responses were manually reviewed for clarity, relevance, mathematical consistency, and overall response quality. Based on these evaluations, the fine-tuned model produced noticeably more relevant and structured responses for the target domain compared to the base model.

Perplexity was also used as a supplementary quantitative metric and showed a modest improvement after fine-tuning. One of the main challenges was the limited amount of high-quality domain-specific Turkish training data. During experimentation, I frequently encountered overfitting, where the model adapted strongly to the training samples without producing equivalent improvements on unseen inputs. This required adjusting the fine-tuning process and carefully evaluating whether improvements actually generalized beyond the training data.

The project gave me practical experience with LoRA-based LLM fine-tuning, human-in-the-loop evaluation, model evaluation, dataset limitations, and identifying and dealing with overfitting in low-data settings.

Stack: Python, PyTorch, Transformers

Currently in development.

A desktop application that works with a browser extension that imports conversations from LLM platforms and processes them locally using embedding models and topic clustering. It groups semantically related discussions, builds an interactive graph of connected topics, and provides an intuitive way to explore previous questions, answers, and context transitions while keeping data on the user's machine.

Stack: Electron.js, llama.cpp, Node.js, Python, FastAPI, HuggingFace

Implemented Gradient Descent, Stochastic Gradient Descent, and ADAM from scratch in C to classify two image classes.

Stack: C

Through this project, I gained hands-on experience with designing and implementing full-stack applications. I developed RESTful APIs, implemented CRUD operations, and built a clean, well-structured frontend and backend architecture. I also learned authentication and authorization with JWT, database design and normalization, and best practices for organizing scalable applications.

Stack: React Native, Node.js, Express.js

Together with my teammates, we developed and released a Turkish-focused model family on Hugging Face under the name Cisimcik AI Lab. This is the name we chose for our independent research team, an informal collective.

Built on Qwen3.5-4B-Base, the family covers continued pre-training, conversational post-training, and decision modeling. All three models are released under the Apache 2.0 license:

• cisimcik-base: a Turkish-first base model trained on 10 billion tokens, available for further fine-tuning.
• cisimcik-4b: a 4.2B-parameter chat model with hybrid reasoning and tool calling, developed through five stages of post-training.
• tau-4b: a System One decision model for yes/no, choice, and scoring tasks, returning calibrated probabilities in a single forward pass with a 32K-token context.

In our published benchmark comparison, cisimcik-4b ranks first on CETVEL with an average score of 57.31 and second on OpenLLM Turkish with 59.60. The model card includes the evaluation results and comparison table.

We are continuing our research and plan to publish an academic paper on our upcoming work.

Stack: Python, PyTorch, Transformers, Hugging Face