claude

ml-engineering

Name: ml-engineering
Author: Matt Corbett

ML-engineering (MLOps) team — agents (ml-platform-architect, training-pipeline-engineer, model-serving-engineer, ml-monitoring-engineer) for the PRODUCTION lifecycle of ML models: the platform/architecture (build-vs-buy, the stack, the train->register->serve->monitor loop), reproducible training pipelines + experiment tracking + a model registry, feature stores and train/serve consistency (avoiding training-serving skew and leakage), model serving (online vs batch, the right deployment with shadow/canary), and monitoring (data + concept drift, performance decay, retraining triggers). skills, a decision-tree knowledge bank (serving-pattern + retraining-trigger trees + a dated 2026 map), best-practices, templates, commands, an advisory hook. Seams: 'is the result significant?' -> applied-statistics, data pipelines -> data-platform/data-streaming, LLM/agent apps -> claude-app-engineering, deploy infra -> devops-cicd/cloud-native-kubernetes. Requires ravenclaude-core@>=0.7.0.

9 on GitHubby Matt Corbettv0.4.6MITupdated today

Plugin ml-engineering from marketplace ravenclaude in mcorbett51090/RavenClaude. Commands read from the repo's manifest — today.

This plugin runs code on your machine — its hooks fire automatically on events like tool use. The exact commands are listed below; read them before installing.

What it installs

Skills (6)

computer-vision-pipeline
Run the computer-vision MLOps lane end to end: data/annotation -> CV task -> architecture choice -> training (transfer-learn first) -> eval by task (mAP/IoU/CER/OKS) -> serving/edge placement. CV-specific leakage (scene-aware split, augment-after-split). Seams back to training/serving/monitoring age
feature-store-consistency
Prevent training-serving skew: compute features once via a feature store or shared transformation so training and serving use identical logic, with point-in-time correctness for temporal features and no leakage of future data.
ml-experiment-tracking
Playbook for setting up and operating an experiment tracking system (MLflow or Weights and Biases) — what to log, run comparison workflow, promotion to the model registry, and avoiding the common leakage and cherry-picking pitfalls.
model-monitoring
Keep production models honest: monitor input/data drift, prediction/concept drift, and performance decay (when labels arrive); define the retraining trigger up front (schedule/threshold/drop); alert on model health; and close the loop to retraining.
model-serving
Serve models reliably: choose online vs batch by the use case, deploy a versioned model from the registry, optimize latency to a budget (batching/quantization/distillation/hardware), and roll out safely with shadow -> canary -> full.
reproducible-training
Build reproducible training: a versioned prep->train->evaluate->register pipeline (not a notebook), experiment tracking (params/metrics/code/data/env), a model registry as source of truth, and leakage-free time-aware validation.

Slash commands (4)

/ml-engineering:build-training-pipeline
Build a reproducible training pipeline with experiment tracking, a registry, feature consistency, and leakage-free validation.
/ml-engineering:deploy-model
Deploy a model: choose online vs batch, serve a registered version, optimize latency, and roll out shadow->canary->full.
/ml-engineering:design-mlops-platform
Design the train->register->serve->monitor MLOps lifecycle matched to team maturity, with a build-vs-buy stack choice.
/ml-engineering:monitor-model
Set up drift + decay monitoring with thresholds, alerts, and a retraining trigger that closes the loop to training.

Subagents (4)

ml-monitoring-engineer
Use for production model monitoring: data drift, prediction/concept drift, performance decay (when labels arrive), defining the retraining trigger (schedule/threshold/drop), distinguishing data vs concept drift, alerting on model health, and closing the loop to retraining.
ml-platform-architect
Use for MLOps architecture: designing the full train->register->serve->monitor lifecycle, build-vs-buy of the ML stack matched to team maturity, structural reproducibility (versioned data/code/config/env + model registry), and governance/lineage.
model-serving-engineer
Use for model serving: choosing online vs batch inference, serving from the registry as a versioned artifact, latency/cost optimization (batching, quantization/distillation, hardware) to a budget, and safe rollout (shadow -> canary -> full) with a promotion metric and rollback.
training-pipeline-engineer
Use for reproducible training: pipelines (prep->train->evaluate->register), experiment tracking, a model registry, feature stores / shared transforms for train-serve consistency, leakage-free time-aware validation, and budgeted hyperparameter tuning.

Hooks (1)

PreToolUse
/usr/bin/env -u BASH_ENV -u ENV -u SHELLOPTS -u BASHOPTS -u LD_PRELOAD -u LD_LIBRARY_PATH -u LD_AUDIT -u DYLD_INSERT_LIBRARIES -u DYLD_LIBRARY_PATH -u PYTHONPATH -u PYTHONSTARTUP PATH=/usr/local/bin:/

Language servers (1)

python

View source on GitHub mcorbett51090.github.io

More from mcorbett51090/RavenClaude

All 184 plugins →

RavenClaude Core

claude

Domain-neutral Claude Code foundation: specialist agents, skills, gates, hooks, rules, templates. Capability Grounding Protocol, Structured Output Protocol, Researcher meta-skill, comfort-posture dashboard, opt-in command-review tribunal (the Thing), WebFetch sanitizer, branch-archive escape hatch, brand-extraction (homepage → brand kit), /wireframe (describe anything → validated model + high-fi Artifact + Mermaid), Problems & Resolutions view, declarative-visualization (Vega-Lite/Deneb/SVG, Gate 101 + Gate 103 svg-report-lint, viz-spec-reviewer), and Claude orchestrator escape hatch (/claude-orchestrate). Slash commands: /init-agent-ready, /wrap, /set-posture, /dashboard, /forge, /wireframe, /reset-plugin-cache, /handoff.

70 skills · 14 commands · 18 agents · hooks

power-platform

claude

Power Platform specialist team — agents (incl. power-platform-tester, power-bi-engineer) and skills, with strong ALM/git coverage for solutions, flows, and PBIP. A house-opinions hook flags 8 §3/§4 violations; the knowledge bank carries production decision trees (PA-flow recovery, Dataverse token-acquisition, PCF React surface, PBI deploy/refresh, custom-connector build-path, PBIR Enhanced infinite-spinner debug + full build reference, DAX silent-zero scoring via the `Domain` pattern, sempy.fabric notebook reference, Power BI Copilot report-readiness, Code Apps connector gotchas, PBIP deployment variables + #839, PBIR reference enrichment, PBIP report fast-solve triage router [MCP-optional], main-forms-are-Web-API-authorable grounding) plus a real-engagement scenarios bank. Bundles the community pbix-mcp server (d0nk3yhm/pbix-mcp, MIT) for .pbix/.pbit read/write/DAX-eval (`pip install pbix-mcp`); documents (not bundles) the official Microsoft Dataverse MCP (CLAUDE.md §9a). Extends ravenclaude-core.

24 skills · 8 commands · 11 agents · 1 MCP server · hooks

finance

claude

Corporate finance & FP&A specialist team — FP&A analyst (budgets, forecasts, variance commentary), financial modeler (3-statement, DCF, scenarios), controller (close, JEs, reconciliations), treasury analyst (cash, covenants, FX), valuation analyst (DCF + comps + precedent), audit-prep specialist (PBC, walkthroughs, SOC1/2 narratives), and board-pack composer. skills incl. a gold-standard NetSuite close (OAuth2 M2M + SuiteQL BS/IS trial balance, layperson runbook), templates, a knowledge bank (ASC 805 / 718 / 740, accrual-cutoff, WACC, cost accounting, FP&A operating model + decision support, NetSuite integration landscape), advisory hooks (hardcoded rates / plaintext PII / missing Sources / Assumptions; secrets-PII scan). Inherits ravenclaude-core protocols.

23 skills · 7 commands · 7 agents · hooks

web-design

claude

Web design & build specialist team — web architect (IA, stack, hosting), UX designer (wireframes, flows, conversion), visual designer (brand, typography, tokens), frontend implementer (HTML/CSS/JS, React, Astro, Next), content strategist (copy, microcopy, SEO content), accessibility auditor (WCAG 2.2 AA/AAA), and performance engineer (Core Web Vitals, image/font optimization). skills (the gold-standard-website-pipeline gated build pipeline, static-site build, design-system audit, card/tile UI, accessibility review, Core Web Vitals tuning, technical SEO audit, information architecture, conversion design, content audit, design-tokens scaffolding, third-party hygiene, Fluent UI v9 + React implementation), templates, an advisory hook, and a knowledge bank seeded with a curated 2026 reference set of "cutting edge yet simple" marketing-site donors — the design agents carry compact inline priors plus a pointer to the full brief. Inherits ravenclaude-core protocols.

19 skills · 7 commands · 7 agents · hooks · LSP

edtech-partner-success

claude

EdTech-flavored Partner Success Manager team — specialist agents anchored on the PSM lane, with skills, templates, an advisory hook, a knowledge bank with Mermaid decision trees, a engagement bank, and a stdlib psm_calc.py calculator (seat-utilization / renewal-NRR-forecast / time-to-value). Covers QBRs, success plans, health scores, a self-contained Power-BI-style partner-health report, onboarding, touchpoint logs, adoption sequencing, 90-day implementation arcs, train-the-trainer curricula, advocacy/reference pipelines, renewal decision memos, and the integration contract for a centralized CS dashboard (Planhat + Salesforce + Snowflake + Zapier + Granola). Team growth ships as skills + knowledge + templates, not parallel agents. Primary-source citations with retrieval dates. Vertical-explicit but segment-agnostic — fits K-12, higher-ed, corporate L&D, or mixed EdTech books. Inherits ravenclaude-core protocols including the alternate-methods Capability Grounding Protocol.

17 skills · 5 commands · 6 agents · hooks

regulatory-compliance

claude

Financial-regulatory & compliance specialist team — function agents (AML/KYC, regulatory-reporting, risk-and-controls, policy & procedure, examination-prep, Bermuda-insurance) PLUS jurisdiction/regulator specialists: BMA financial-institutions (banking, trust, corporate-services, fund-admin, investment-business, money-service, digital-assets — the primary build-out, with deep AML/ATF + supervision/fees + licensing-classification trees), CIMA (Cayman), Bahamas, Channel Islands (Jersey JFSC + Guernsey GFSC), UK PRA, and US federal+state. primary-source-cited regulator knowledge files, best-practice rules, skills, templates, a defensive PreToolUse hook (PII scrub — SSN/IBAN/credit-card/Bermuda TIN/wire instructions). Positioning reflects field experience inside a Tier-1 financial regulator (BMA). Inherits ravenclaude-core protocols.

11 skills · 5 commands · 12 agents · hooks

Similar plugins

trackio-cli

claude

Query Trackio projects, runs, and metrics using the `trackio` CLI. Use when the user needs to list projects/runs/metrics, get summaries, or retrieve metric values from local Trackio databases. Supports both human-readable and JSON output formats. Covers project discovery, run inspection, metric querying, and system metrics. Designed for LLM agents and automation scripts.

1 skill

sickn33/agentic-awesome-skills 49,620

dominodatalab

claude

Full Domino Data Lab platform support — workspaces, jobs, model deployment, experiment tracking, GenAI tracing, Spark/Ray/Dask, and app deployment for data science teams

26 skills · 4 commands · 3 agents · 1 MCP server

anthropics/claude-plugins-official 47,076

machine-learning-ops

claudecodex

ML model training pipelines, hyperparameter tuning, model deployment automation, experiment tracking, and MLOps workflows

2 skills · 1 command · 3 agents

wshobson/agents 42,461

deep-learning-book

claude

Study companion for the Deep Learning textbook by Goodfellow, Bengio & Courville (MIT Press, 2016), free to read at deeplearningbook.org. Twenty chapter files, a glossary, patterns and a cheatsheet index the whole book, and a delta reference dates it against 2026 practice with per-claim confidence levels: double descent qualifying the U-curve, AdamW splitting weight decay from L2, transformers displacing Chapter 10's recurrence, diffusion growing out of Chapter 18's score matching. Four stdlib tools make it executable — a prerequisite-closed reading-path planner that refuses goals the book does not cover, a training diagnostic running Chapter 11's rules in priority order so a NaN is never reported as overfitting, a capacity planner that ranks 'shrink the model' last when overparameterized, and a parameter/FLOP/activation-memory calculator that refuses a stack whose shapes do not connect. A companion, not a compilation: the book is copyrighted, so nothing here reproduces its text.

1 skill · 3 commands · 1 agent

alirezarezvani/claude-skills 29,725

voltagent-data-ai

claude

Data engineering, ML, and AI specialists - data pipelines, machine learning, LLM architecture

VoltAgent/awesome-claude-code-subagents 26,513

agents-data-ai

claude

Agents for data engineering, machine learning, and AI development

12 agents

davepoon/buildwithclaude 3,741

Maintain ml-engineering?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins184 plugins