AI Engineer at b(x) Theory
May 2026 – Aug 2026Built the document intelligence and financial RAG platform behind a restructuring analytics product
Adam Kamel
Software Engineering University of Waterloo
I'm a Software Engineering student at UWaterloo working across ML research and applied AI. I've published on interpretability and context compression, and I build reinforcement learning agents, including a Minecraft combat bot that beats 99% of human players. Right now I'm co-leading WorldFold at Wat.AI, teaching a dual-arm robot to fold cloth with world models. Open to new opportunities and collaborations.
Experience
Technical Project Manager at Wat.AIDesign Team
May 2026 – PresentLeading WorldFold, a project exploring world models and reinforcement learning for robotic cloth manipulation
Machine Learning Engineer at WatStreetDesign Team
Dec 2025 – Aug 2026Trained statistical and deep learning classifiers to identify market regimes in noisy financial time series
Software Engineering Intern at Workbench
May 2025 – Sep 2025Developed AI agent integrations and improved natural language processing systems
Papers
Emergent World Beliefs: Exploring Transformers in Stochastic Games
Adam Kamel et al.
NeurIPS 2025 MechInterp Workshop · EMNLP 2025 BlackboxNLP Workshop
Explores whether transformers build internal world models of games they can't fully observe. Generated 3M+ simulated poker games to train a GPT-2-style model on hand data, then probed its internal activations, recovering hand strength with 98% accuracy and showing the model encodes its own win probability (r = 0.59), information it was never shown during training.
RSCE: Training-Free Residual Stream Encoding for Persistent Context Amortization
Adam Kamel, Eric Xu
KnowFM Workshop @ ACL 2026
A training-free method that compresses long documents into reusable vector representations. Benchmarked against leading compression baselines on QA and code tasks across 7B–70B models, RSCE retains 60–81% of answer quality while using 99% fewer tokens, and the paper identifies why quality drops when it does.
Featured Projects
Minecraft RL Combat Agent
Self-play RL agent that beats 99% of human players in Minecraft sword duels
- C++
- PyTorch
- Java
- Reinforcement Learning
Pufferfish
Chess engine that placed 3rd overall out of 75 teams at ChessHacks
- Python
- Pytorch
- Huggingface