Portfolio · plain reading view
I build machine learning systems, research software, and practical tools across AI, retrieval, backend systems, and Linux.
Computer Science graduate with concurrent bachelor's degrees from Blekinge Institute of Technology and Andhra University. I work across machine learning, research software, backend systems, local AI, information retrieval, and Linux. My BTH thesis on personalized federated learning for StarCraft II became a first-author IEEE GEM 2026 paper. I am currently a Linux Kernel Summer Mentee and a research collaborator on ongoing game-AI work.
Machine Learning & Software Engineer | Researcher
Machine Learning & Software Engineer | Researcher · tarunb.co@gmail.com · tarunb.me
Thesis: Federated Learning-Based Player Personalization in Real-Time Strategy Games: Player Action Prediction in StarCraft II. The thesis work became a first-author paper presented at IEEE GEM 2026.
GPA 8.65/10 · WES iGPA 4.0/4.0 · Concurrent dual-degree program.
Contributing to three ongoing game-AI research projects with different research teams. Work includes dataset analysis, benchmark development, research software, experiments, evaluation, and reproducible workflows. Some project details remain non-public.
Selected for the Linux Kernel Summer Mentorship program. Working with upstream Linux development, subsystem analysis, debugging, builds, and patch workflows using C, Git, GCC, Kbuild, QEMU, GDB, and Bash.
Delivered weekly one-to-one web-development tutoring. Designed project-based exercises and coached debugging and independent problem solving in JavaScript, HTML/CSS, and responsive design.
Built and validated LightGBM/XGBoost models, feature-engineering and cross-validation pipelines, Flask/FastAPI inference APIs, and ROC-AUC/SHAP analyses for client-facing ML work. Worked with relational databases and presented model behavior and results to stakeholders.
Prepared meeting minutes, reports, and summaries of internal and external discussions.
Built Flask applications backed by SQLite/MySQL plus Python tools for file handling and visualization, including a CSV-to-histogram web app, and deployed projects on PythonAnywhere.
Computer Science graduate working on personalized and federated learning, player modeling, sequential data, research software, local AI, information retrieval, and machine learning systems. My BTH thesis became a first-author IEEE GEM 2026 paper, and I am currently collaborating on additional game-AI research.
This work studies next-action prediction from professional StarCraft II replay data under centralized and simulated federated training. The benchmark uses the last 8 events of a replay window, 56 input features, GRU/LSTM/Transformer models, and 387 exact action labels. The processed scope contains 10,088,820 action events from 11,552 replays; 1,305 training-set player identifiers act as candidate federated clients. I compared centralized training, FedAvg, FedProx, and a personalized shared-backbone/local-head design across 30 trained configurations. Centralized models reached roughly 41–44% Exact Top-1 accuracy across races, while the best federated results were about 34.7–39.1%, a gap of roughly 4–7 percentage points. Transformers were strongest under federated training in the reported experiments. The backbone/local-head runs demonstrated local parameter separation, but did not establish a positive personalization gain. Reported configurations use one fixed seed, and the federated setup is a simulation of data-local training rather than a deployment with differential privacy or secure aggregation.
First-author paper, presented September 2026.
Collaborating with Dr. Prashant Goswami and other researchers across three ongoing game-AI projects. Current work includes dataset analysis, benchmark development, research software, experiment tooling, evaluation, and reproducible workflows; some project details remain non-public.
Research software and reproducibility artifacts for the StarCraft II federated-learning thesis and IEEE GEM 2026 paper. Includes preprocessing, centralized and federated training, GRU/LSTM/Transformer models, evaluation, checkpointing, communication accounting, saved protocol metadata, tests, figures, and citation files.
Calibrated linear-SVM NLP pipeline using word/character TF-IDF and engineered cue features. Current held-out evaluation reports 0.9378 macro-F1 on 123,905 examples.
Local-first information-retrieval system with document parsing, source locators, hybrid vector/FTS5 retrieval, local Ollama inference, incremental indexing, and evidence-bearing outputs.
Computer-vision research/software prototype for pose comparison with MediaPipe and OpenCV, supporting offline analysis, live feedback, rendered comparison videos, and joint-level analytics.
Research software and reproducibility artifacts for my first-author IEEE GEM 2026 work on personalized federated learning for StarCraft II player-action prediction.
Local-first folder intelligence system that indexes documents and code, preserves source locations, and answers questions through hybrid vector and keyword retrieval with local Ollama inference.
Semantic novel search and personal reading library with local embeddings, FAISS, lexical boosts, SQLite persistence, ratings, collections, reading status, and tier lists.
Computer-vision prototype for teacher/student pose comparison with offline and live webcam workflows, joint-level error analytics, rendered comparison videos, CLI tooling, and a Streamlit interface.
Calibrated linear-SVM classifier for questions, tasks, and deadlines using word/character TF-IDF, cue features, lexical features, reusable inference helpers, CLI tooling, and a Streamlit demo.
Team-built Industrial AI Hackathon prototype connecting supplier risk to contract rights, deadlines, evidence, and suggested actions, with deterministic offline operation and optional live tool-calling workflows.
Research software for my first-author IEEE GEM 2026 work on personalized federated learning for StarCraft II player-action prediction.
Python · PyTorch · Federated Learning · GRU/LSTM/Transformer
Processed 10,088,820 action events from 11,552 replays with 387 exact action labels. Compared centralized training, FedAvg, FedProx, and a personalized shared-backbone/local-head design across 30 trained configurations. The repository includes preprocessing, training, evaluation, checkpointing, communication accounting, tests, result aggregation, figures, citation metadata, and reproduction instructions.
Local-first folder intelligence prototype for research, code, documents, and other collections.
Python · Ollama · Chroma · SQLite FTS5 · RAG
Parses PDF, EPUB, DOCX, text/code, CSV, and XLSX while preserving source locations. Uses hybrid vector and FTS5 retrieval with reciprocal-rank fusion, local workspace/chat persistence, incremental content hashing, bounded exhaustive workflows, source-grounded outputs, and chapter-aware spoiler filters. The current backend keeps inference local through Ollama.
Semantic novel search and personal reading library.
FastAPI · FAISS · sentence-transformers · SQLite · JavaScript · Docker
Combines local sentence embeddings with exact/token lexical boosts and tag filters. Supports personal libraries, reading status, ratings, CSV imports, S-D tier organization, catalogue/admin workflows, CI, Docker deployment, backup tooling, and index/catalogue consistency checks.
Multilingual museum ticket booking application built with Flask and Python, using Google Maps and Google Translate APIs for booking and interactive maps.
Personal training prototype using computer vision with MediaPipe and OpenCV.
Custom blog application built with Flask and SQLite/MySQL.
Python web application for file handling and CSV data visualization as histograms.
Contributing to three ongoing game-AI research projects with different research teams. Work includes dataset analysis, benchmark development, research software, experiment tooling, evaluation, and reproducible workflows; some project details remain non-public.
Selected for the Linux Kernel Summer Mentorship program. Working with upstream Linux development, subsystem analysis, debugging, build and patch workflows using C, Git, GCC, Kbuild, QEMU, GDB, and Bash.
Delivered weekly one-to-one web-development tutoring, designed project-based exercises, and coached students through debugging and problem solving in JavaScript, HTML/CSS, and responsive design.
Built and validated LightGBM and XGBoost models, feature-engineering and cross-validation pipelines, and Flask/FastAPI inference APIs backed by relational databases. Produced ROC-AUC and SHAP analyses and presented model behavior for client-facing work.
Prepared meeting minutes, reports, and concise summaries for internal and external use.
Built Flask applications backed by SQLite/MySQL, plus Python tools for file handling and visualization including a CSV-to-histogram web application; deployed projects on PythonAnywhere.
Four year dual degree program in collaboration with BTH.
Started a four year bachelor's degree program, finally getting a bachelor in technology in computer science engineering with honors from Andhra University, and a Bachelor of Science in Computer Science specializing in Artificial Intelligence and Machine Learning from Blekinge Tekniska Högskola.
Built a Flask blog application backed by SQLite/MySQL and Python web tools for file handling and data visualization, including a CSV-to-histogram web app.
Prepared meeting minutes, reports, and summaries of discussions.
Built and evaluated LightGBM and XGBoost pipelines, worked on feature engineering and cross-validation, exposed inference through Flask and FastAPI APIs, and used ROC-AUC and SHAP analyses for client-facing machine-learning work.
Delivered weekly one-to-one web-development tutoring, designed project-based exercises, and supported debugging and problem solving.
Selected for the Linux Kernel Summer Mentorship program and working with upstream Linux development, debugging, builds, Git workflows, GCC, Kbuild, QEMU, GDB, Bash, and C.
Collaborating on dataset analysis, benchmark development, experiment tooling, and reproducible research workflows.
Selected for the program; working with upstream Linux development.
Presented first-author work on personalized federated learning for StarCraft II player-action prediction.
Certificate of completion.
Certificate of completion.
PY0101EN, provided by IBM.
RP0101EN, provided by IBM.
Mobile SEO course with Greg Gifford.
tarunb.me