The library

constitutional-ai · When you want safety alignment without human labels and need explainable reasoning in refusals.

diff-since-my-commit · Code review workflows where you need to see what others changed to your files.

llamaguard · Production input/output filtering where you need a specialized 7B moderation model instead of general LLM.

pubfi-dsl-server-contract · When you need strict input validation and reject SQL/unknown-field queries server-side.

nemo-guardrails · When you need stateful conversational guardrails beyond single-turn content moderation.

gitlab-protected-branches · Multi-team workflows where you must enforce review gates and prevent force-push accidents.

prompt-guard · When you need lightweight client-side jailbreak detection before sending to LLM.

huggingface-accelerate · When you have a standard PyTorch training loop and need easy multi-GPU scaling.

augmented-reality-tracking · Diagnosing why virtual AR content does not register with physical world; guides technique selection.

training-llms-megatron · Training LLMs over 1B parameters where single GPU is insufficient; achieves 47% MFU on H100.

pytorch-fsdp2 · Sharding large models across GPUs with DTensor-based parameter sharding and simpler checkpoint semantics vs FSDP1.

code-review · Post-implementation review to catch security issues, bugs, and code quality gaps before merge.

pytorch-lightning · Scaling training from laptop to multi-node/multi-GPU without rewriting boilerplate; automatic DDP/FSDP/DeepSpeed support.

spatial-reasoning-fundamentals · Grounding 3D spatial thinking in platforms like Minecraft, VR, CAD; makes blueprint→world translation transferable.

ray-train · Running hyperparameter searches or long training runs across unstable/cloud clusters with automatic fault recovery.

lambda-labs-gpu-cloud · Quick provisioning of single or batch GPU instances without long-term commitment; lower cost than AWS on-demand.

rust-build · Building production Rust binaries with optimizations (LTO, codegen-units) for deployment.

beads-state · Building reactive UIs in Beads where state changes automatically propagate to views.

modal-serverless-gpu · Deploying inference-only ML models without managing containers or servers; pay-per-invocation.

skypilot-multi-cloud-orchestration · Running large training jobs by automatically selecting cheapest cloud provider and handling spot preemption.

awq-quantization · Compressing large LLMs for edge deployment while maintaining near-FP16 accuracy.

mail-async · Multi-agent orchestration where messages must survive crashes and maintain order by timestamp

quantizing-models-bitsandbytes · Fitting 7B+ models on consumer GPUs (8-16GB VRAM) when accuracy tolerance permits <1% degradation

nudge-sync · Gastown health checks and stall detection where only the latest signal matters and fast overwrite is critical

optimizing-attention-flash · Long-context LLM inference and training where attention is the bottleneck and 2-4x speedup with 10-20x memory savings is critical

migrate-to-shoehorn · Test suites with large mock objects where only a few properties matter and traditional `as` is heavy-handed

token-budget · Multi-agent Gastown convoys needing hard token gates to prevent runaway spend and cascade failures

gguf-quantization · Deploying LLMs on consumer hardware (MacBook M1+) or servers without NVIDIA GPU when universal hardware support is required

gptq · Deploying 70B+ models on A100/H100 when 4× compression and <2% accuracy loss is acceptable

k-fencore · WoW addon authors building complex addon logic without reimplementing common utilities

hqq-quantization · Fast model quantization when calibration data unavailable and extreme compression (2-bit) is acceptable

webapp-testing · End-to-end regression testing when unit tests miss user-facing behavior

ml-training-recipes · Starting model training quickly with expert-vetted defaults instead of tuning from scratch

commands-gizix-cc-projects · Extending Claude Code project templates with custom agent behaviors

evaluating-code-models · Comparing code model performance across standard benchmarks when new architecture or training method is evaluated

init-pablolion-bmad-plugin · Starting a new project using BMAD methodology with pre-built scaffolding

evaluating-llms-harness · Standardized model comparison using industry-standard benchmarks when you need reproducible academic metrics.

api-design · Creating APIs that agents and humans consume reliably when predictability beats cleverness.

nemo-evaluator-sdk · Enterprise benchmarking of multiple models at scale when reproducible containerized evaluation is required.

llama-cpp · Edge deployment and local inference on Apple Silicon and non-NVIDIA hardware without Docker complexity.

detecting-command-injection · Finding command injection during code review when automated scanners miss indirect and complex paths.

sglang · Agentic workflows with repeated prefixes (system prompts, tools) where 5× speedup via caching outweighs setup.

code-health-check · Regular codebase audits to catch health degradation before it blocks features.

serving-llms-vllm · Production serving of open models on NVIDIA hardware when high throughput and predictable latency required.

mlflow · Managing dozens of experiments with repeatable comparison and governance when you need audit trails.

idiomatic-rust · Systems code where memory safety must hold at compile-time without runtime GC overhead.

experiment-tracking-swanlab · Real-time experiment monitoring during training when live curves beat post-hoc analysis.

tensorboard · When debugging deep learning models with real-time metric visualization and experiment comparison across runs.

weights-and-biases · When managing production ML experiments that need team collaboration, automatic metric logging, and hyperparameter optimization.

commit · When creating git commits that reflect user intent and preserve reasoning for future context.