Research-oriented software developer working across computational mathematics, AI systems, GPU computing, formal methods, and reproducible scientific software.

My public repositories are primarily research prototypes and experimental systems. I try to keep claims narrow, preserve machine-readable evidence where practical, and separate exploratory results from publication-grade conclusions.

Computer-assisted graph-invariant research centered on exact finite-order classification and reconstruction-style questions. The current public evidence line studies aligned vertex-deleted invariant families on the complete order-10 graph catalog, with explicit claim boundaries and reproducibility artifacts.

Experimental GPU acceleration for a selected Lean/CIC-style type-checking fragment. The project is scoped as systems research rather than a replacement for the Lean kernel, and keeps correctness checks and benchmark evidence alongside performance work.

Experiments in lowering a supported Python subset into tensor-oriented execution paths, with correctness checks and raw benchmark evidence.

• ToposAI — experimental neuro-symbolic research library combining finite categorical/topos-inspired structure with PyTorch components.
• paradigm-stack — parameter-matched architecture experiments with repeated-seed benchmark evidence.
• algebraic-pruning — structured pruning experiments with explicit evaluation scope and retained evidence.
• cr-molecular-fingerprint — exploratory molecular representation research; results should be read with the repository's stated baseline and leakage limitations.

• coherence-lab — certificate-producing normalization and equivalence experiments for selected categorical coherence fragments.
• proof-perf-lab — performance experiments paired with explicit correctness checks and reproducible reporting.
• contextdiet — tooling for auditing coding-agent context such as AGENTS.md, CLAUDE.md, skills, and MCP configuration.
• universal-ai-dev-assistant — experimental developer-assistant platform and integration workspace.

• gpu-assembly-index — GPU-oriented assembly/indexing research with documented provenance.
• gpu-shap — experimental GPU attribution tooling with an intentionally limited scope.

• Scope before headlines. Finite experiments, prototypes, and restricted fragments should be described as such.
• Claims need provenance. Important results should identify code, data, configuration, environment, and evidence artifacts.
• Optimization needs correctness. Performance results are useful only when the compared behavior is checked.
• Negative results are part of the record. Failed hypotheses, counterexamples, and corrected claims should remain visible.
• Reproducibility beats promotional numbers. Raw or machine-readable evidence is preferred over isolated benchmark claims.

• exact finite computational mathematics and graph invariants
• GPU-accelerated formal and symbolic computation
• efficient AI inference and model-system experiments
• reproducible benchmark design and scientific software
• research automation with explicit evidence and claim boundaries

Some repositories are historical experiments, snapshots, or archived research branches. The projects listed above are the clearest public entry points into the current portfolio.