[Yaseen Khalil]|Computational Modeler & ML Systems Architect
> Exploring the mathematical architecture of intelligent systems. Bridging high-dimensional feature engineering with production data pipelines and autonomous AI integrations.
Technical Matrix
- 01PCA / SVD
- 02Multivariate linear regression
- 03Lasso (L1) feature selection
- 04Ridge (L2) multicollinearity control
- 05Madelon regularization under distractors
- 01BiLSTM + attention anomaly detection
- 02PyTorch sequence models
- 0310-minute micro-batch training
- 04DBSCAN hotspot clustering
- 05Sequential anomaly pipelines
- 01Semi-Tensor Product (STP) algebraic linearization
- 02Attractor dynamics
- 03Boolean network dynamics
- 04Intervention scoring (pyMaBoSS)
- 01Python
- 02Java
- 03R
- 04SQL
- 01TypeScript
- 02Go
- 03Mojo
- 04KDB/q
- 05CSS
- 06PostgreSQL
Systems Architecture
> Full-stack mobile/web cosmetologist marketplace. Architected robust user auth and multi-party payment routing (barber payouts). Optimized with AI-assisted code generation.
> Conceptual AI agent gateway. Unified system interactions for Slack, Google Workspace, Square, and Sentry APIs.
> Efficiency-Fidelity benchmark comparing Semi-Tensor Product (STP) algebraic linearization (exact logic preservation) vs multivariate least-squares regression (scalable dynamics) for cancer signaling circuits. Implementing both linearizations against stochastic Boolean networks with pyMaBoSS, starting from gene-expression driven circuits to evaluate behavior under drug-like perturbations. Developing intervention scoring approaches, transitioning from rule-curated Boolean subnets (~50-200 nodes) toward larger PPI graph structures.
> Production-style Airflow 3 pipeline ingests partitioned Levin vehicle telemetry, normalizes fields into a canonical schema, and generates curated daily rollups. Enforced idempotent loads (safe reruns/backfills via unique event keys), dynamic task mapping for date-range backfills, and run-level artifacts for operational visibility. Data contract validations: schema/type/range checks (timestamp parseability, non-null vehicle IDs, plausible bounds for speed/RPM/temp) with quality stats per partition.
> End-to-end vehicle health neural network using bi-directional LSTM with attention; improved anomaly detection accuracy from 40-60% to 87-95% through iterative retraining on 30K+ telemetry points every 10 minutes. Three-tier anomaly detection: LSTM for sequential insights, DBSCAN for geographic hotspot clustering; 33% increase in predictive reliability across simulated fleets. FastAPI microservices on Railway with Supabase integration.
Blog
Engineering a Cell: From 17,000 Dimensions to a Single Matrix