About
I'm Thabolezwe Mabandla, a graduate student in Engineering Artificial Intelligence at Carnegie Mellon University Africa. My work sits at the intersection of applied AI, software engineering, robotics, and language technology.
I have built production-grade data pipelines, RAG systems, and LLM-powered applications for cybersecurity, media, and enterprise AI use cases. I'm especially interested in systems that combine robust machine learning infrastructure with real-world deployment constraints.
My current research direction focuses on adversarial reinforcement learning and robotics, while my broader interests include world models, computer vision, vision-language models, drone technologies, and African language technologies.
Education
Aug 2025 - May 2027 Master of Science in Engineering Artificial Intelligence - Carnegie Mellon University Africa
Graduate program focused on Applied AI, with growing specialization in embedded systems, robotics, and intelligent autonomous systems.
- Concentration: Applied AI
- Applied AI
- Robotics
- Embedded Systems
Aug 2020 - Jun 2024 BSc Honours in Artificial Intelligence & Machine Learning - University of Zimbabwe
Completed undergraduate research on low-resource NLP for Zimbabwean languages, including a Shona spell checker for language digitization and preservation.
- Concentration: Natural Language Processing
- Thesis: Digitization of Indigenous Zimbabwean Languages: A Shona Spell Checker
- Graduated with First Class Honors
- Awarded the University of Zimbabwe Book Prize
- Awarded Best Undergraduate Researcher and Innovator in the Faculty of Computer Engineering
- NLP
- Machine Learning
- Low-resource Languages
Experience
Jun - Aug 2026 Graduate Research Assistant - CMU-Africa
Conducting research on adversarial reinforcement learning and robotics with Professor Jesse Thornburg, focusing on robust learning methods for embodied intelligent systems.
- Adversarial RL
- Robotics
- Reinforcement Learning
- Research
Mar - Aug 2025 Backend Developer / Machine Learning Engineer - CyberAlliance
Contributed to SallyAI, a threat intelligence platform delivering real-time cybersecurity insights through LLM and code agents. Engineered distributed data pipelines and deployed retrieval and threat-intelligence infrastructure for enterprise security workflows.
- Built pipelines with Apache Airflow, Kafka, Redis, and Postgres.
- Fine-tuned LLMs with HuggingFace Transformers, BitsandBytes, MLflow, and AWS.
- Implemented RAG with Weaviate, Elastic Stack, and Sentence Transformers.
- Integrated CTI sources including NVD NIST, Abuse.ch, AlienVault OTX, and MITRE.org.
- Python
- LLMs
- Airflow
- Kafka
- Redis
- Postgres
- Weaviate
- AWS
Oct 2024 - Mar 2025 AI Software Developer - Rubiem Innovations
Led development of local LLM applications and enterprise AI prototypes, including an agent for Zimpapers that helped journalists retrieve internal knowledge and search the web during article writing.
- Designed LLM applications for customer support and content generation.
- Supported AI strategy work for banks and telecom clients.
- Translated technical concepts into client-facing recommendations and training.
- LLM Agents
- RAG
- Google Search
- Client Delivery
- AI Strategy
Jan - Jun 2024 Research Scientist - Zimbabwe Centre for High Performance Computing
Conducted NLP research for indigenous Zimbabwean language technology and developed a novel Shona spelling correction model on high-performance computing infrastructure.
- NLP
- HPC
- Shona
- Language Technology
- Model Training
Sep 2022 - Aug 2023 IT Operations Intern - Old Mutual & CABS
Supported infrastructure operations for financial systems by monitoring uptime, improving deployments, maintaining documentation, and executing system maintenance and compliance checks.
- IT Operations
- Infrastructure
- Documentation
- Security Compliance
Research Projects
Sep 2026 - Present Agent-Augmented VLA for Training-Free Policy Adaptation
Designing an agentic supervision harness that monitors a frozen vision-language-action policy at runtime, detects execution failures, and applies corrective actions without retraining or collecting new demonstrations. Evaluating the approach on LIBERO manipulation tasks before sim-to-real transfer to bimanual YAM robots.
- Vision-Language-Action
- Agentic AI
- LIBERO
- Robot Manipulation
- Sim-to-Real
May 2026 - Present Adversarial RL for Indirect Robot Control
Developed a ROS 2 and Gazebo reinforcement learning framework in which an embodied parasite agent steers an autonomous iRobot Create 2 indirectly through virtual bumper interventions. Modeling temporally extended actions as an SMDP and training PPO policies for sparse, goal-directed control under partial observability.
- Reinforcement Learning
- ROS 2
- Gazebo
- PPO
- SMDP
- Robotics
Aug 2023 - Jun 2024 LLM for Shona Spelling Correction
Developed the first LLM-based spell checker for Shona, supporting language digitization and preservation for one of Zimbabwe's major native languages. Built and fine-tuned models on HPC clusters as part of undergraduate research.
- LLMs
- Shona
- NLP
- HPC
- Research