Internship at IBM Research explores tools for mechanistic interpretability

Curiosity about the intersection of machine learning and AI led Baimam Boukar Jean Jacques to a summer internship at IBM Research in Kenya. Baimam is pursuing his master of science in information technology at Carnegie Mellon University Africa (CMU-Africa).

“One of the biggest takeaways from my internship is getting to understand mechanistic interpretability—one of the most active research areas in AI,” he said. Boukar’s internship at IBM Research focused on IBM Deep Scanner, a tool used previously to analyze how large language models represent toxic input prompts or personas internally.

He chose to study at CMU-Africa because, “CMU-Africa graduates shine at the international level and the university offers a global environment with world-class professors,” he said.

Boukar credits what he learned at CMU-Africa and his experience as a software engineer as instrumental to his internship, because it required extensive software knowledge to work as an architect on IBM Deep Scanner.

CMU-Africa graduates shine at the international level and the university offers a global environment with world-class professors.

Baimam Boukar Jean Jacques, student, CMU-Africa

During his internship, he created a new modular architecture for IBM Deep Scanner that allowed him to integrate multiple legacy codebases into a new architecture. He also enhanced the process of running experiments in cluster environments, enabling scientists to conduct them more quickly and efficiently.

“The internship focused on how we can build tools that will enable us to better understand how large models are working internally and how they are predicting outcomes.” he explained. “Imagine deploying AI in high-stakes scenarios like aerospace systems and it crashes the whole system, possibly endangering lives. That is not acceptable and with mechanistic interpretability we can develop methodologies to mitigate those issues.”

The team also wrote a research paper, “Retrieval with Multiple Query Vectors Through Anomalous Pattern Detection,” which has been accepted for a presentation at the workshop “New frontiers in information retrieval” held at the 40^th Conference on Artificial Intelligence of the Association for the Advancement of Artificial Intelligence. The conference will be held in Singapore in January 2026. The team also has a patent related to their work that was recently approved.

During his internship, Boukar joined mentors and fellow interns in activities, including ice skating, hiking, playing ping-pong, going to a football game, and playing games with other scientists in the lab.

After graduation this spring, Boukar plans to pursue his doctorate or explore industrial AI research opportunities that combine his interests in industrial AI, machine learning, and aerospace. “The internship also gave me the opportunity to work with experts and a network of scientists.”  Boukar said. “I learned a lot from them on how to structure ideas, run research experiments, and write research papers. I believe these are instrumental skills.”