Verapulse
OngoingI’m building a platform for accessing real robots and collecting robot data remotely.
Hi, I’m
Burak Barlas.
I build software and work
on robot learning.
I’m a CS master’s student at USC.
Outside of work, I sail and play chess.
Selected work
Building & experimenting Learning to pick and place
ExperimentTraining an SO-101 policy in simulation, then turning it into a dataset of demonstrations.
VLM4Rec
Multimodal Semantic Representation for Recommendation with Large Vision-Language Models.
Research details
LVLM-Based Multimodal Representation Learning for Recommendation
- Converted product images into semantic text with LLaVA-NeXT-7B, then encoded the text with Sentence-BERT for retrieval.
- Improved Recall@10 by 54.9% across 4,700+ items.
- Benchmarked several multimodal architectures; the text-only LVLM approach outperformed attention and spectral fusion methods.
- Built an offline pipeline using 4-bit quantized LVLM inference for 4,700+ items.
More projects & research
GLAMOR Lab · USC
Vision-Language-Action Representation Analysis
Studied what a robot policy’s internal representations reveal about its behavior.
- Probed π0-FAST to extract vision, language, and action representations on LIBERO manipulation tasks.
- Tested language and visual perturbations, finding failures caused by reworded instructions and texture changes.
- Found that representation similarity did not reliably predict policy success under these perturbations.
Text-to-image fine-tuning
Multimodal Figure Generator
Fine-tuned Stable Diffusion to generate academic figures from LaTeX captions.
- Used LoRA, Hugging Face, and PyTorch for fine-tuning.
- Built a dataset by extracting and aligning figure-caption pairs from arXiv papers in LaTeX and PDF formats.
- Designed a prompt-refinement pipeline to improve diagram clarity and text legibility.
Senior design research project
RL-Based Home Energy Management
Used model-free Q-learning to manage household energy use under changing electricity prices.
- Trained a Q-table model and predicted household energy usage in five-minute intervals.
- Reported 87% forecasting accuracy and 17% energy cost savings compared with the algorithms evaluated in the project.
C++ · 3D graphics
3D Spatial Rendering and Interaction Pipeline
Built a basic rendering pipeline to explore how objects and cameras behave in a 3D scene.
- Implemented geometric primitives, transformation matrices, and perspective projection in C++.
- Explored how spatial transformations affect object alignment and visual consistency.
- Worked with coordinate systems, camera perspective, and scene composition as foundations for immersive systems.
C++ · Information retrieval
Efficient Search with NLP
Built a search algorithm using TF-IDF, cosine similarity, and synonym matching.
- Preprocessed text with stop-word filtering, tokenization, and stemming.
- Combined those steps with cosine similarity and synonym matching to improve precision and contextual relevance.
Recent writing
Notes from what I’m learningMarkov Decision Processes: Decisions, Made Visible
An interactive guide to how rewards and actions shape an agent’s decisions.
Teaching an SO-101 to Pick and Place
Notes on training a pick-and-place policy and turning it into 1,000 demonstrations.
How Do We Compare Vision-Language-Action Models?
Comparing models by task success, instruction following, and changes in the scene.
Experience
Researcher
LIRA Lab · USC
Sep 2026 — now
Course Producer · DSML / DS 552
University of Southern California
Sep 2026 — now
Founder
Verapulse
Jun 2026 — nowPrevious experience
Researcher
GLAMOR Lab · USC
May — Aug 2026
Founding Engineer
Leap
Mar — Aug 2025Software Developer
Private company
Sep 2024 — Mar 2025
Software Engineer
Vagon.io
Aug 2022 — Aug 2024Software Developer Intern
May — Aug 2022Software Developer Intern
May — Aug 2021Education
M.S. Computer Science
University of Southern California · Los Angeles, US
Aug 2025 — May 2027 · expectedCoursework
- Applied Natural Language Processing
- Deep Learning and Its Applications
- Machine Learning
- Algorithms
B.S. Computer Science
Bahçeşehir University · Istanbul, Turkey
Sep 2018 — Jun 2023Coursework
- Introduction to Machine Learning
- Introduction to Neural Networks
- Data Structures and Algorithms (C++)
- Digital System Design
- Object Oriented Programming (C++)
- Database Management Systems
Tools I work with
- Languages
- C++ · Swift · Python · JavaScript · TypeScript
- Machine learning & robotics
- NVIDIA Isaac Sim · PyTorch · CoreML · Hugging Face · LoRA · LIBERO
- Apps & infrastructure
- SwiftUI · UIKit · React Native · React · Next.js · AWS · Firebase