A ML playground for education purpose

Fullstack machine learning inference template

A systems programming language where automatic differentiation is a compiler pass and model parameters are explicit, growable memory.

A Branchless, Zero-Jitter Ingress Router for 32-GPU Distributed Mesh Networks utilizing JAX/XLA and NCCL.

tpu-doc is a zero-dependency diagnostic binary for Google Cloud TPU environments that instantly validates hardware health, discovers software stack configurations, and provides AI-powered log analysis to eliminate expensive debugging downtime.

A custom Deep Learning Framework built from scratch using NumPy. Implements Automatic Differentiation, Backpropagation, and modular Layer architectures.

Discrete-event simulation benchmark for LLM request cancellation handling, measuring wasted compute, KV residency overhead, and survivor latency recovery across polling and push-based strategies.

Analytical benchmark connecting pipeline parallelism and KV cache management: bubble-time overlap for disaggregation, prefix reuse tax of PP stages, and tiering speedup across NVLink, InfiniBand, and PCIe

Compiler MVP that detects Transformer fusion patterns, generates optimized CUDA kernels with WMMA Tensor Cores, and executes them on real GPU hardware — 10.5 TFLOPs on RTX 2070, correctness validated against PyTorch.

Discrete-event simulation benchmark for graceful degradation strategies in LLM serving: output capping, CoT disabling, model downgrade, and selective drop under progressive pressure.

🔍 Explore machine learning hands-on with scikit-learn through projects that illustrate key concepts like classification, regression, and clustering.

A research-grade course in Numerical Methods and Optimization for ML. No deep learning libraries—just pure math and NumPy. Rebuilds automatic differentiation engines, stable SVD/Cholesky algorithms, quasi-Newton solvers (BFGS), stable log-sum-exp layers, and Gaussian Process regressions from scratch

Tungsten Alpha: A JAX-native, 7-layer Adelic-Riemannian Operating System. Implements resurgent gradient flow through discrete logic and Levi-Civita parallel transport for decentralized XLA clusters. Hardened for zero-jitter, systolic array execution.

Simulation benchmark for canary deployment strategies in LLM model updates, measuring safety, rollback behavior, bad exposure, and rollout speed across good, bad, and marginal candidate scenarios

Distributed, GPU-aware workload scheduler for heterogeneous clusters: queueing, quotas, GPU flavors, and autoscaling.

Analytical benchmark for sliding window attention KV cache management: quality vs window size tradeoffs, SWA vs eviction comparison, prefix sharing interaction, and operational window recommendations across four attention distributions

Monte Carlo simulation benchmark for MoE expert activation memory variability and admission control under non-deterministic routing — including routing skew amplification, top_k sweep, and policy comparison across uniform, skewed, and very_skewed distributions

Discrete-event simulation for LLM serving capacity planning: how many GPUs for a p99 TTFT SLO under ShareGPT traffic, batching fragmentation, and autoscaling lag — with cliff analysis and three levels of planning rigor.

Calibrated simulation benchmark for real-time LLM request routing, comparing complexity signals, output-length awareness, cost savings, and quality-risk trade-offs.

Discrete-event simulation benchmark for decode preemption policies in LLM serving: when to interrupt a running decode to admit an urgent prefill, and whether checkpoint or recompute is cheaper.