Become an inference infra engineer by mid-2029.

A verifiable CKA certificate ID or official result record, with the completed private road-engine journal and evidence at cka-vim-first-principles/ referenced as supporting proof; the credential remains supplemental to repository-level engineering evidence.

A verifiable CKS certificate ID or official result record saved in the private evidence directory, with the associated security-lab notes or repository link; retain this credential proof alongside, not instead of, the repositories, traces, and reproducible measurements in the other tracks.

A verifiable CKAD certificate ID or official result record saved in the private evidence directory; retain this credential proof alongside, not instead of, the repositories, traces, and reproducible measurements in the other tracks.

A verifiable KCNA certificate ID or official result record saved in the private evidence directory; retain this credential proof alongside, not instead of, the repositories, traces, and reproducible measurements in the other tracks.

A verifiable KCSA certificate ID or official result record saved in the private evidence directory; retain this credential proof alongside, not instead of, the repositories, traces, and reproducible measurements in the other tracks.

A repository-linked record naming the chosen hypervisor and why; the VM definitions or exact launch commands; free and df output from inside each VM; and measured host disk consumed by the substrate. The record must show the 78 GB free-disk constraint and whether the suggested or reduced sizing was used.

The exact kubeadm configuration used; an unedited transcript containing kubectl get nodes -o wide for the control plane and workers; the containerd setup; the chosen CNI with its reason; and notes on every bootstrap failure and repair.

A full command transcript from etcd snapshot creation through deliberate cluster-state destruction and restore, including snapshot metadata or checksum, an inventory of what was lost versus recovered, and a measured restore duration supported by timer output or start and end timestamps.

A before-and-after version record for Kubernetes, kubeadm, kubelet, and kubectl on every node, the ordered procedure actually followed with the exact commands and drain and uncordon output for each worker, and a written account of anything that broke with its repair.

A per-drill record for at least five distinct failures, including kubelet down, expired certificate, misconfigured static pod, CNI failure, and disk-pressure eviction; every record must state the first symptom, hypothesis, first broken layer identified, minimal repair, and verification output.

A node inventory with verified CPU, RAM, disk, architecture, NIC, and KVM readiness for homelab1, homelab2, and workstation; the physical network topology and MTU; the exact kubeadm join procedure across physical hosts; evidence of the LoadBalancer/MetalLB or equivalent and real-storage behavior; and an explicit written comparison of what differed from the Stage 1 VM build, including a genuine node or link failure.

A versioned repository containing a CMake-built Tensor2D<float> implementation with unit tests, Debug and Release builds, sanitizer output, NumPy or PyTorch correctness comparisons, and benchmark output across multiple matrix sizes.

A versioned repository plus benchmark report comparing naive, loop-reordered, tiled, SIMD, and multithreaded matmul; the report names hardware and compiler, records runtime and GFLOP/s at multiple sizes, shows thread scaling and correctness, and explains the bottleneck behind each change.

A repository-linked study with raw perf stat and perf record/report output plus an annotated assembly or flamegraph artifact for the matmul workload, including command lines, hardware and software context, baseline-versus-optimized measurements, and a causal explanation.

A versioned infrastructure repository with manifests or Helm values, successful GPU Operator installation evidence, kubectl output for driver, device-plugin, and DCGM pods, and a reproducible workload log proving a scheduled GPU allocation.

A versioned deployment repository and reproducible load-test report showing the model endpoint scaling on GPU-utilization or queue/request metrics, with workload details, replica transitions, throughput, TTFT, TPOT, and p99 latency recorded.

A versioned dashboard definition and exported view that joins request, GPU-utilization, token-count, and accelerator-cost data to show throughput, TTFT, TPOT, p99, utilization, and cost per output token, with query and reproduction instructions.

A versioned CI workflow with model checksum or immutable version checks, correctness and smoke tests, reproducible image or package build, staging deployment evidence, and a demonstrated rollback or promotion record without embedded secrets.

A versioned CUDA repository containing vector-add, reduction, matmul, softmax, and RMSNorm kernels with correctness tests across multiple shapes, warmup and repeated measurements, and profiler traces that classify each result as compute-, memory-, synchronization-, or launch-bound.

A repository-linked study using a named model and hardware that measures accuracy delta against a reference, memory footprint, and throughput change across FP8 and INT4 weight/activation quantization, with AWQ/GPTQ, paged attention, KV-cache management, and chunked prefill configurations; include reproducible commands.

A versioned Triton repository with reference-based correctness tests, a shape and dtype sweep covering batch and sequence variation, benchmark tables with named hardware and software versions, an exported profiler trace, and documented loss cases.

A reviewable commit or pull-request URL in a vLLM or SGLang fork that changes a real runtime path, with passing correctness or regression tests, a reproducible benchmark command, and a short explanation of the affected mechanism.

A repository-linked before-and-after load-test report naming hardware, model, workload, concurrency, warmup and repetition method, and exact commands; it must report TTFT, TPOT, output-token throughput, request throughput, p99 latency, memory use, and the causal bottleneck analysis.

A repository-linked study with single- and multi-GPU measurements, an explicit scaling-efficiency calculation, per-rank NCCL or Nsight Systems traces, throughput and communication-overhead tables, hardware topology context, and reproducible commands.

A SUBMITTED upstream pull request URL with genuine review engagement: a documented maintainer response, review comments addressed, CI passing, and a reproducible performance or correctness result documented; a later merge may be recorded as a bonus but is not required for unlock.