Projects
Building the infrastructure for autonomous, self-improving AI—from multi-agent orchestration to safe synthetic data pipelines.
Physics-Based AI Image Detection
ActiveDetecting AI-generated images through physics-based reasoning — analyzing depth maps, brightness-depth consistency, and light estimation to expose how AI fails to model real-world physics. A 3-feature classifier achieves 68.3% accuracy with just depth gradients and brightness edge analysis.
- •Key finding: AI images are paradoxically too physically consistent — overly smooth brightness-depth relationships and symmetric depth distributions
- •Yet AI images show sharper depth gradients (d=0.653), revealing lack of true 3D understanding
- •3-feature model (grad_mean, brightness_at_depth_edges, n_valid_patches) outperforms full 27-feature model by 13.3pp
- •27 physics features across depth statistics, brightness-depth coupling, and light estimation pipelines
- •PCA reveals real/fake signal is not dominant — scene complexity dominates, requiring targeted feature selection
DREAM-C2L: Continual Learning Framework
ActiveOpen-source framework for continual learning research. Enabling AI systems to learn continuously without catastrophic forgetting, adapting to new data while preserving prior knowledge.
- •Difficulty-aware sample ordering algorithms
- •Replay-based and regularization methods for knowledge retention
- •Reproducible experiment pipelines for HPC clusters
- •Integration with PyTorch Lightning and Weights & Biases
Project Canary
CompletedFoundational MOVE Fellowship project (Sept-Oct 2025) — a community-driven effort to train and refine frontier AI models. Completed 15,000+ tasks across 15 domains, improving Review 1 approval rates from 10% to 40%.
- •High-volume task generation across CS, Math, Medicine, Physics domains
- •Core contributor in Computer Science domain
- •Quality improvement: raised approval rates from 10% to 40%
- •Precursor to Project Orion's specialized refinement phase
Project Orion
CompletedAdvanced MOVE Fellowship phase (Nov 2025) — specialized refinement of frontier AI models. Focused on high-quality reasoning chains, safety injections, and red-teaming through jailbreak testing. One-month intensive following Project Canary.
- •High-quality reasoning refinement and chain-of-thought improvement
- •Safety injection tasks: embedding guardrails into model behavior
- •Red-teaming and jailbreak testing for frontier models
- •Built on Canary foundations with deeper specialization in CS domain
The Future of AI (2026+)
The industry is pivoting from building oracles (models that talk) to building partners (systems that act).
Agentic AI Systems
Moving from 'answer my question' to 'do this for me.' Multi-agent ecosystems where specialized models collaborate—one plans, another executes, a third verifies.
Insight: Users need autonomy. AI that can browse, book, coordinate, and handle tasks without constant supervision.
Test-Time Compute & Reasoning
Scaling inference over training. Models that 'think longer' before responding, allocating compute dynamically to solve complex logic and reasoning problems.
Insight: Users need reliability. No more confidently stated hallucinations—systems that check their own work.
World Models & Physical AI
Training AI on video and sensor data to understand cause-and-effect. Critical for the robotics surge and real-world AI applications.
Insight: Users need contextual awareness. AI that can 'see' and understand physical environments, not just parse text.
Small, Specialized & Sovereign
Edge AI and Mixture of Experts (MoE) models that run locally on phones and laptops. Cheaper, faster, and more accurate for specialized domains.
Insight: Users need privacy and speed. Local AI without sending sensitive data to distant cloud servers.
Interested in Collaboration?
Open to research partnerships, consulting, and investment opportunities.
Get in Touch