Ali Torkamani

About Ali Torkamani

I work on improving the reliability of AI systems.At Amazon Web Services, I focus on improving the logical reasoning capabilities of language models using reinforcement learning, structured reasoning signals, and verifiable feedback.The goal is to move beyond outputs that appear plausible toward systems that are logically consistent and reliable in real-world settings.This builds on a broader theme across my work: learning from imperfect signals. This perspective is rooted in earlier work on robust and adversarial machine learning.Previously, in AWS Security, I worked on large-scale machine learning systems for threat detection, including Mithra, a graph-based system operating on billions of nodes and edges for detecting malicious domains. I also worked on foundation models for AWS entity representation learning and developed Firenze, a framework for evaluating models without reliable ground truth.Earlier, I worked in healthcare, building machine learning systems for patient modeling, clinical decision support, and predictive analytics.Across these areas, the underlying question has been:How can systems learn effectively when supervision is incomplete, noisy, or unreliable?I am also the author of “AI Engineering Interviews” (O’Reilly).

Experience

1. Senior Applied Scientist Aws Ai

   Amazon Web Services (AWS)

   Dec 2024 — Present · New York, NY, US

   Leading research and development on improving the logical reasoning capabilities of large and small language models through advances in training, post-training, and model architecture- Designing post-training approaches that combine reinforcement learning with structured reasoning feedback to improve logical consistency and factual correctness- Exploring learning paradigms based on verifiable and structured signals, moving beyond reliance on human preference alone- Developing methods for training and adapting mixture-of-experts (MoE) and hierarchical transformer architectures to improve efficiency and specialization- Investigating pre-training strategies and curriculum learning techniques to improve reasoning capabilities and data efficiency- Building systems that integrate generation, validation, and iterative refinement to increase model reliability- Partnering with engineering teams to productionize models, optimizing for latency, cost, and real-world deployment constraints

Education

• University of Oregon

  Doctor of Philosophy (PhD), Computer Science

  2011 — 2016

• Isfahan University of Technology

  Master of Science (M.Sc.), Artificial Intelligence

  2003 — 2006

• Oregon State University

  Doctor of Philosophy - PhD, Machine Learning and Computer Vision

  2009 — 2011

Skills

• Python
• Latex
• Algorithm Design
• Artificial Intelligence
• Apache Spark
• Software Project Management
• Scipy
• High Performance Computing
• Machine Learning
• Opencv
• Algorithms
• Data Analysis
• Neural Networks
• Recommender Systems
• Information Retrieval
• Signal Processing
• Writing
• Statistics
• Mysql
• Distributed Systems
• Sql
• Linux
• Eclipse
• Computer Science
• C
• Programming
• Optimization
• Gurobi
• R
• Software Engineering
• Cplex
• Java
• C++
• Data Structures
• Graph Theory
• Mathematical Modeling
• Computer Vision
• Natural Language Processing
• Scala
• Game Theory