Illustration: How alloy microstructure changes during phase separation.
See my work on
detecting such transitions without labels.
abram.mj@gmail.com Google Scholar GitHub LinkedIn CV
I'm a theoretical physicist, ML engineer, AI researcher, and systems architect. I work at FirstPrinciples, a Canadian AI lab, building agentic systems to advance fundamental research. I specialize in multi-agent systems, LLM reliability, and scientific applications of both classical ML and agentic AI. Currently, I am building an evaluation harness for our scientific reasoning platform and investigating failure modes in LLM-assisted generation.
Research on how frontier language models and agents fail on expert-level scientific work, how we can detect those failures, and how to build more reliable agentic systems.
Built an evaluation harness that verifies the reasoning and output of agentic scientific systems, with symbolic, numerical, and conceptual checks.
Led roadmap planning for the research unit and assumed the role of architect for non-platform delivery, including verification as a service.
Research on distinct failure modes of enterprise GenAI systems. Co-developed IBM's roadmap for next-gen agentic AI development.
Built several PoCs and MVPs for enterprise and government clients. In one of the engagements, proposed an agentic retrieval system that cut hallucinations by over 80%.
Led the Data and AI group in my unit as technical lead. Led technical discussions with client executives. Organized both internal and client-facing workshops.
Research on scientific machine learning and the role of noise in the quality of LLM-assisted generation.
Built the curriculum of a new graduate program that combines physics and AI.
Led the Machine Learning for the Natural Sciences group. Supervised three PhD theses and eight capstone projects.
Research on self-supervised detection of microstructural transitions in materials, and on privacy-preserving federated training that is robust to data poisoning.
Built an MVP for Sandia National Laboratories, assisting with the development of a system for informing the synthesis of new functional materials.
Led the creation of the Machine Learning for the Natural Sciences group. Taught undergraduate physics and graduate machine learning classes.
Research on novel consensus algorithms and the role of machine learning in decentralized multi-agent systems. Investigated novel attack vectors and the role of architecture in mitigating their impact.
Built test suites, benchmarks, and attack simulations to evaluate the performance of the consensus protocol.
Led the redesign of the main consensus protocol. Technical lead through a series of pivots. Led the development of a patented method for collective learning in decentralized systems.
Research on developing embeddings from highly structured data and applying NLP-assisted classification in highly regulated industries.
Built an end-to-end machine learning pipeline (from data collection to training) as a founding engineer in a lean team, writing a large part of the startup's codebase. Our results briefly established a new SOTA in the field of credit risk assessment for the SME market in the UK.
Led the data collection, model training, and model evaluation effort.
Research on novel approximate methods for modeling superconducting and magnetic quantum systems.
Built high-performance simulation code in C++.
Led an outreach program for gifted high-school students preparing for the physics olympiad. Taught statistical physics classes.