Cagan Bakirci
MS in Computer Science (AI) · University of Southern California
I am a Master's student in Computer Science at USC. I'm fortunate to be affiliated with LIRA Lab and CPS-VIDA Lab, where I work with Prof. Erdem Bıyık and Prof. Jyotirmoy V. Deshmukh on making robot learning more reliable for real-world deployment. My research interests include safety-constrained reinforcement learning and uncertainty quantification for vision-language-action policies.
Research
Safe Reinforcement Learning for Robotic Manipulation
Developing methods to improve safety in learned robotic manipulation policies while maintaining task performance across benchmark environments.
LIRA Lab · with Prof. Erdem Bıyık
Failure Detection for Vision-Language-Action Policies
Investigating uncertainty quantification techniques for detecting potential failures in manipulation policies conditioned on language instructions.
CPS-VIDA Lab · with Prof. Jyotirmoy V. Deshmukh
Learning-based Coordination for Multi-Agent Pathfinding
Exploring learned reward functions for improving coordination in decentralized multi-agent navigation scenarios.
CPS-VIDA Lab · with Prof. Jyotirmoy V. Deshmukh
Selected Projects
Projects from graduate coursework and undergraduate research
EEG-to-Text Decoding with LLM Rescoring
Optimized EEG-to-text decoding via beam search tuning and no-repeat n-gram constraints, improving non-teacher-forced BLEU-1 by 32% (relative) on ZuCo V1.
CSCI 544 · Fall 2025
Reconstructing Music from EEG Responses
Reconstructed music stimuli from EEG by mapping PSD features to mel-spectrograms; achieved 82.86% classifier accuracy and 0.80 cosine similarity, surpassing EEG2Mel baseline.
CSCI 566 · Fall 2024
Earthquake Time-to-Failure Prediction
Developed neural network models (CNN, RNN) to predict earthquake time-to-failure from seismic data. Achieved top 5% ranking in LANL Kaggle competition with strong generalization.
CMPS 242 · Spring 2019 · Kaggle Top 5%
Software & Other Projects
Artsy: Artist & Artwork Search App
Full-stack web application for searching artists and artworks via the Artsy API. Features include favorites, detailed artwork views, and responsive design.
CSCI 571 · Fall 2024
Amazon Alexa Modular Adapter
Contributed to an Amazon-sponsored accessibility device enabling voice control of switch-operated appliances for users with disabilities. Developed voice command integration using Alexa Voice Service SDK and Alexa Skills Kit.
UCSC Capstone · 2020 · Amazon-Sponsored
Industry Experience
- Promoted to Senior in two years (fastest timeline permitted by company policy)
- ComMind: Founded and led development of a dynamic, self-correcting rating system with ML models. Gained recognition from Sabanci University and secured placement in Koç University's incubation hub.
- Search Helper: Co-developed scalable microservice for query correction, automatically fixing 210,000+ typos monthly and boosting search reliability by 89%.
- Oculus: Led computer vision system to automate reading of stock/serial numbers on products.
- DropShipment: Architected and led implementation of fulfillment module, enhancing operational efficiency and service delivery.
- Backend Development: Implemented features on SAP Hybris Commerce using Java and Spring Boot.
- Developed asynchronous, reactive microservices using Eclipse Vert.x in Java 15, with PostgreSQL for data storage, improving system reliability.
Education
Contact
Best reach: caganbakirci [at] gmail [dot] com · USC: cbakirci [at] usc [dot] edu