I architect high-efficiency AI systems by fusing Python’s ML velocity with Rust’s runtime performance.
From neuromorphic edge inference to graph-scale intelligence, I optimize for latency, throughput, and cloud cost.
I operate at the intersection of HPC (High-Performance Computing) and Applied AI.
My engineering philosophy is simple: I don’t just train models; I bend them into production constraints—compressing complexity, removing runtime bottlenecks, and maximizing efficiency per compute dollar.
My core specialization is building Hybrid Runtimes: Python for rapid ML iteration, Rust (via PyO3) for memory-safe, concurrent, low-latency execution.
This eliminates GIL-bound bottlenecks and enables predictable scaling under real production load.
- AI Runtime Optimization: Rust-driven execution paths for real-time inference and data preprocessing.
- Graph + Neuromorphic Intelligence: GNNs for topological reasoning, SNNs for temporal edge AI.
- Extreme Compute Efficiency: Memory reduction, CPU scaling, and elimination of Python's execution overhead.
- Imbalance & Heterophily Mastery: Loss-level mathematical tuning for asymmetric real-world datasets.
graph LR
A[Python Prototyping] --> B[Profiling Bottlenecks]
B --> C[Bending Layer Optimization]
C --> D[Rust Runtime / PyO3]
D --> E[Optimized Engine]
E --> F[Stable Production]
A high-performance engine combining GNN heuristics with Ant Colony Optimization.
Impact: Migrated critical search loops to Rust, achieving thread-safe concurrency and removing GIL overhead.
SNN architecture (184-input → 120-LIF) with PostPre STDP feature extraction.
Impact: Engineered for edge-native medical IoT. Achieved sub-5ms inference target with ~97% memory reduction.
Advanced GNN pipeline (GCN/GAT) for Bitcoin transaction monitoring.
Impact: Solved severe class imbalance and heterophily using mathematically tuned weighted objectives.
| Project | Stack | Performance Signal | Engineering Impact |
|---|---|---|---|
| GNN_Guided_ACO | PyTorch + Rust | Near-instant convergence | Hybrid-Parallel Execution |
| SNN-Seizure | BindsNET + STDP | < 5ms Inference | ~97% Memory Reduction |
| Elliptic-Fraud | PyG + XGBoost | 96.35% F1-Weighted | Robustness under imbalance |
- < 5ms target latency in edge-native SNN inference.
- 30× throughput scaling potential by moving critical bottlenecks from Python to Rust.
- O(V + E·H) complexity awareness for stable graph-scale production planning.