A comprehensive list of my software development, AI, and machine learning projects.
Turing AI OS (Developed on a 14-year-old PC - 2012)
Engineered Turing AI OS, an AI-native Linux desktop layer integrating local LLM inference, agentic system tools, persistent memory, and natural-language OS interaction using Python and KDE Neon.
Integrated Ollama and LangChain to run privacy-first local LLM inference with token streaming, configurable models, temperature controls, and a lightweight qwen2.5:1.5b default model optimized for CPU-based execution.
Implemented persistent AI memory using ChromaDB, embedding and storing conversations locally.
Developed agentic system capabilities including a natural-language terminal that converts user intent into executable Bash commands, file-system operations, and context-aware code/project analysis with user confirmation.
Built a modular PyQt6 desktop interface featuring an always-available AI Sidebar, Spotlight command palette, Vision analyzer, and AI Control Panel, creating an integrated AI-native operating environment with low-resource local execution.
Developed an end-to-end ML and data science platform enabling dataset upload, automated data cleaning, EDA, and visualization for
supervised and unsupervised learning tasks.
Implemented training pipelines for regression, classification, and clustering algorithms, with intelligent algorithm suggestions based on target
variables.
Enabled model export functionality allowing users to download trained models for reuse in external applications.
Integrated OpenRouter API for visual LLM to find hidden patterns of images, extending platform capabilities beyond tabular data.
Built a no-code, user-friendly interface using Streamlit, enabling non-technical users to perform end-to- end ML workflows.
Developed FABX, an AI-powered reverse engineering and file intelligence platform that analyzes executables, ML models, databases, and structured files, transforming complex software artifacts into structured technical intelligence.
Engineered a multi-format analysis pipeline supporting EXE, DLL, ELF, SO, BIN, ONNX, PyTorch, Pickle/Joblib, SQLite, JSON, YAML, INI, and TOML, with automatic file-type detection and specialized analyzers.
Integrated Rizin and local Ollama LLMs to combine traditional static analysis with AI-powered explanations, behavioral analysis, security insights, and best-effort C/C++ source reconstruction from compiled artifacts.
Built automated ML-model intelligence capabilities that extract layers, operators, tensors, parameter counts, inputs/outputs, checkpoint metadata, and generate ONNX/PyTorch architecture visualizations.
Developed an interactive analysis interface using NiceGUI, Plotly, and NetworkX, providing visual graphs for binary structures, model computation graphs, database relationships, and configuration trees alongside AI-generated technical reports.
Oxygen - Multi Agentic AI Framework that works like a virtual tiny company.
Engineered Turing AI OS, an AI-native Linux desktop layer integrating local LLM inference, agentic system tools, persistent memory, and natural-language OS interaction using Python and KDE Neon.
Integrated Ollama and LangChain to run privacy-first local LLM inference with token streaming, configurable models, temperature controls, and a lightweight qwen2.5:1.5b default model optimized for CPU-based execution.
Implemented persistent AI memory using ChromaDB, embedding and storing conversations locally.
Developed agentic system capabilities including a natural-language terminal that converts user intent into executable Bash commands, file-system operations, and context-aware code/project analysis with user confirmation.
Built a modular PyQt6 desktop interface featuring an always-available AI Sidebar, Spotlight command palette, Vision analyzer, and AI Control Panel, creating an integrated AI-native operating environment with low-resource local execution.
Developed NGIBS, a privacy-focused AI research and browsing system that runs fully locally on user hardware.
Implemented multiple AI intelligence modes including Quick Search, Live Search, Deep Search, and Context Mode to support offline
knowledge retrieval, reasoning, and memory-aware conversations.
Designed a modular architecture allowing users to download and integrate any compatible LLM for local inference.
Integrated Ollama as the backend engine to manage local model execution and inference.
Built and trained a lightweight, 4.8 million parameter Optical Character Recognition (OCR) model from scratch on Google Colab using a
CRNN (CNN + BiLSTM) architecture.
Created a memory efficient data streaming pipeline to train the model on 2.32 million images, loading files one by one to completely
bypass hardware and memory limits without downloading the entire dataset at once.
Executed a 3-phase training strategy, starting with a massive synthetic dataset and then fine-tuning the model on real-world scene text to
adapt it to messy, blurry, real-life images.
Achieved 79.20% accuracy on unseen MJSynth images and 78.37% on IIIT5k, maintaining a low Character Error Rate (CER) of 8.56%
and 3.40 CTC loss.
Published the final 18.3 MB model as a public PyPI package (pip install nano-ocr), making it instantly deployable for edge devices,
embedded systems, and low-power hardware.
Developed a deep learning–based PPE object detection system to automatically identify safety equipment (Helmet, Vest, Gloves, Boots) in industrial environments.
Fine-tuned a pretrained YOLOv8 model using transfer learning on a custom Roboflow dataset and trained it on Google Colab.
Evaluated model performance using IoU-based metrics: Precision (89%), Recall (89%), mAP@50 (91%).
Built an AI-powered web app using Gradio and OpenCV for real-time camera-based compliance detection.
Built DocsMind, a RAG-based document assistant that lets users chat naturally with PDF, TXT, and DOCX files.
Developed the complete document-to-answer pipeline using LangChain, ChromaDB, and Ollama, including chunking, embeddings, retrieval, and response generation.
Added source-grounded responses that show the document context used for each answer, improving transparency and reducing hallucinations.
Implemented conversational memory to support follow-up questions and maintain context throughout a document chat session.
Built an LLM-as-a-Judge evaluation system, achieving 5/5 Context Relevance and 5/5 Faithfulness on sample benchmark queries.
Built NLP Intelligent Lab, an interactive Python-based playground for learning and experimenting with NLP and language understanding.
Implemented hands-on visualizations for core NLP concepts including tokenization, stemming, lemmatization, POS tagging, NER, and text preprocessing.
Added practical experiments covering TF-IDF, Bag-of-Words, word embeddings, text classification, and language modeling.
Explored modern deep-learning approaches including RNNs, LSTMs, Transformers, and attention-based language models.
Designed the lab to turn complex NLP concepts into interactive experiments and visual outputs, making model behavior easier to understand and evaluate.
Project Taste - for Developers, Engineers, Tester and ML Engineers
Built Project Taste, a developer-focused toolkit bringing together 150+ practical tools for everyday software development and engineering workflows.
Developed utilities across areas including API testing, JSON/XML processing, encoding/decoding, hashing, regex, networking, and data conversion.
Designed a single, organized interface that reduces the need to switch between multiple online developer tools and improves day-to-day development productivity.
Focused on fast, lightweight, and reusable utilities, making common developer tasks accessible through simple, purpose-built workflows.
Organized and maintained 150+ tools as a growing open-source project, with a focus on usability, extensibility, and practical developer needs.