AI Architecture Implementation: From Design to Production

Guiar la implementación de arquitecturas AI desde el diseño hasta producción — selección de tecnología, implementación de pipelines, configuración de serving, despliegue de monitoreo, y automatización CI/CD. Produce blueprints de implementación, guías de selección de tecnología, y un playbook fase-a-fase que transforma decisiones arquitectónicas en infraestructura operativa.

Principio Rector

1. Phased delivery, not big bang. Implementar en fases con valor entregable por fase. Fase 0 (Foundation) → Fase 1 (Data Pipeline) → Fase 2 (Model Development) → Fase 3 (Serving) → Fase 4 (CI/CD) → Fase 5 (Monitoring). Cada fase produce capacidad usable.

2. Start simple, evolve with evidence. No sobre-ingenierar para escala hipotética. Empezar con la implementación más simple que resuelva el problema actual. Feature Store no es necesario para un modelo; multi-model tiering no es necesario para un tier.

3. Tests and monitoring from Phase 0, not Phase 5. La infraestructura de testing y monitoreo se establece en la primera fase, no se agrega después de incidentes. Cada componente implementado incluye sus tests y sus métricas desde el día uno.

Inputs

Parámetros:
  MODO:        [greenfield | brownfield | remediation | migration]
  FORMATO:     [ejecutivo | técnico | híbrido]
  VARIANTE:    [pipeline | serving | genai | mlops | full]
  STACK:       [python | java | typescript | polyglot]
  INFRA:       [kubernetes | serverless | containers | hybrid]

Detección automática:
  - Si existe codebase → MODO=brownfield
  - Si existe audit report → MODO=remediation
  - Si el input menciona "migrar" → MODO=migration
  - Si existe LangChain/LlamaIndex → VARIANTE incluye genai
  - Default: MODO=greenfield, VARIANTE=full, STACK=python, INFRA=containers

When to Use

• Implementar una arquitectura AI diseñada (output de ai-software-architecture o ai-pipeline-architecture)
• Seleccionar stack tecnológico para un nuevo sistema AI (frameworks, tools, databases)
• Implementar Feature Store, Model Registry, o Drift Detection desde cero
• Configurar CI/CD para ML (Blue & Gold deployment, validation gates)
• Desplegar sistema de monitoring y observabilidad para AI
• Implementar pipeline RAG completo (ingestion, chunking, embedding, retrieval, generation)
• Ejecutar roadmap de remediación de una auditoría de arquitectura AI
• Migrar de notebooks a producción (MLOps maturity progression)

When NOT to Use

• Diseñar la arquitectura (antes de implementar) → metodologia-ai-software-architecture
• Auditar un sistema existente → metodologia-ai-architecture-audit
• Seleccionar patrones de diseño → metodologia-ai-design-patterns
• Definir estrategia de testing (sin implementar) → metodologia-ai-testing-strategy
• Implementar específicamente en AWS → metodologia-aws-architecture-implementation
• Implementar infraestructura cloud genérica → metodologia-infrastructure-architecture

Delivery Structure: 6 Sections

S1: Technology Stack Selection

Selecciona las tecnologías para cada componente del sistema AI con justificación basada en constraints.

Load references:
  Read ${CLAUDE_SKILL_DIR}/references/technology-selection.md

Decision Framework:

1. Constraints: Cloud provider, Kubernetes availability, team expertise, budget, regulations
2. Workload type: Traditional ML, Deep Learning, LLM Application, Agentic AI
3. Scale: Prototype, Production single-model, Production multi-model, Enterprise

Selection matrix por componente:

Entregable: Stack decision document con ADRs por cada selección no-obvia.

S2: Data Pipeline Implementation

Implementa el pipeline de datos desde ingestion hasta feature serving.

Load references:
  Read ${CLAUDE_SKILL_DIR}/references/implementation-blueprints.md

Componentes a implementar:

• Data ingestion (sources → raw store)
• Data quality gates (schema validation, distribution checks, business rules)
• Feature engineering (transformation logic, shared between training/serving)
• Feature store setup (if applicable — offline + online + registry)
• Data versioning (training datasets)
• Pipeline orchestration (DAGs, scheduling, retry logic)
• Data monitoring (drift, freshness, coverage)

Blueprint selection:

• Single model, simple features → direct pipeline, no feature store
• Multiple models, shared features → Feature Store (Blueprint 1)
• Streaming requirements → Kafka/Kinesis + streaming feature computation
• Batch only → Airflow/Dagster + batch transformation

Entregable: Implemented pipeline with tests, monitoring, and documentation.

S3: Model Development & Registry Setup

Implementa el ciclo de desarrollo de modelos con tracking, registro, y evaluación.

Componentes a implementar:

• Experiment tracking setup (MLflow/W&B)
• Training pipeline (reproducible: code + data + config = model)
• Evaluation framework (accuracy, fairness, robustness, explainability)
• Model registry (artifact storage, metadata, versioning, staging)
• Hyperparameter optimization (Optuna/Ray Tune integration)
• Fairness testing suite (demographic parity, equal opportunity)
• Explainability integration (SHAP/LIME for traditional ML, attention for transformers)

Blueprint selection:

• MVP → MLflow local, manual training, basic evaluation
• Production → MLflow/W&B + automated training pipeline + full evaluation suite
• Enterprise → Full MLOps platform + governance workflows + compliance

Entregable: Training pipeline, experiment tracking, model registry operational.

S4: Serving & Inference Implementation

Implementa model serving, API layer, caching, y fallback mechanisms.

Componentes a implementar:

• Model serving deployment (vLLM/Triton/TorchServe/Seldon)
• API layer (REST/gRPC endpoints, versioning, documentation)
• Authentication and authorization (API keys, OAuth, service mesh)
• Rate limiting (per-client, per-endpoint)
• Caching layer (prediction cache + semantic cache for GenAI)
• Circuit breaker (fallback cascade: model → cache → previous → denial)
• Load testing baseline (P50, P95, P99, max throughput)

GenAI-specific implementation:

• Guardrails system (input validation, output filtering, PII masking)
• RAG pipeline (Blueprint 5: ingestion, chunking, embedding, retrieval, generation)
• Agent framework (tool definitions, governance, memory, orchestration)
• Multi-model routing (Blueprint 6: tier assignment, cost tracking)

Entregable: Serving infrastructure operational, API documented, load tested.

S5: CI/CD & Deployment Automation

Implementa Blue & Gold deployment con validation gates automatizados.

Componentes a implementar:

• Code CI (linting, type checking, unit tests, security scan)
• Data CI (data quality tests triggered on pipeline changes)
• Model CI (model quality tests triggered on model changes)
• Blue & Gold pipeline (Blue=production, Gold=staging with validation)
• Canary deployment (gradual traffic shift with metric monitoring)
• Rollback automation (one-click rollback to previous version)
• Promotion workflow (automated gates + configurable manual approval)

Gate configuration:

• Code quality gate: lint + type check + unit tests + security scan
• Data quality gate: schema + distribution + freshness
• Model quality gate: accuracy ≥ threshold + fairness ≥ threshold + latency ≤ SLA
• Regression gate: no accuracy drop > X% vs. current production
• Security gate: vulnerability scan + access control verification

Entregable: CI/CD pipeline operational, gates configured, rollback tested.

S6: Monitoring & Observability Implementation

Implementa el stack de observabilidad completo para el sistema AI.

Componentes a implementar:

• Infrastructure monitoring (CPU, GPU, memory, storage, network)
• Application monitoring (request rate, error rate, latency)
• Model monitoring (accuracy tracking, prediction distribution, confidence)
• Data monitoring (input drift, feature freshness, quality violations)
• Drift detection system (Blueprint 3: statistical monitoring + alerts)
• Dashboards (executive, operations, ML, data)
• Alerting (thresholds, escalation policies, on-call integration)
• Runbooks (incident response for common AI failures)

Dashboard hierarchy:

• Executive: Business KPIs, model impact, cost summary
• Operations: System health, latency, errors, throughput
• ML: Model metrics, drift, fairness, prediction quality
• Data: Pipeline health, quality scores, freshness

Entregable: Full observability stack, dashboards, alerts, runbooks.

Trade-off Matrix

Assumptions

1. Arquitectura AI ya diseñada (output de ai-software-architecture, ai-pipeline-architecture, o similar)
2. Cloud provider y presupuesto definidos
3. Equipo técnico disponible con habilidades core (Python, ML basics, CI/CD)
4. Ambiente de desarrollo provisionado o provisionable
5. Acceso a datos de entrenamiento (o plan para obtenerlos)

Limits

1. NO diseña la arquitectura — implementa una arquitectura ya diseñada
2. NO selecciona patrones — implementa patrones ya seleccionados (ver metodologia-ai-design-patterns)
3. NO genera modelos ML — implementa la infraestructura para entrenarlos y servirlos
4. Implementación AWS-específica requiere metodologia-aws-architecture-implementation
5. NO reemplaza la expertise del equipo — guía y estructura, no ejecuta en solitario

Edge Cases

1. Equipo sin experiencia ML: Fase 0 extendida con capacitación. Empezar con managed services (SageMaker, Vertex). Reducir complejidad de Feature Store y multi-model tiering hasta que el equipo madure.

2. Migración desde notebooks: Priorizar extracción de feature engineering y training logic a módulos Python testeables. Notebooks quedan solo para exploración. Fase 2 se convierte en la fase más larga.

3. Remediación post-auditoría: Ordenar implementación por priority score del audit report, no por la secuencia estándar de fases. Puede requerir empezar por Fase 5 (monitoring) si el hallazgo crítico es "no hay observabilidad".

4. Sistema GenAI puro (sin ML tradicional): Fases 2 y 3 se fusionan en "RAG/Agent Implementation". Feature Store no aplica. Focus en guardrails, vector DB, prompt management, cost controls.

5. Restricciones de presupuesto extremas: Open-source everything. MLflow (free), Feast (free), Evidently (free), GitHub Actions (free tier). Docker Compose para desarrollo, single-instance para producción inicial.

Validation Gate

• Stack tecnológico seleccionado con ADR por cada decisión no-obvia
• Repository structure creada con tests y CI desde Fase 0
• Data pipeline implementado con quality gates operativos
• Feature store/pipeline produciendo features consistentes (training = serving)
• Model registry operativo con versioning, metadata, y staging workflow
• Serving layer operativo con API documentada y load tested
• Guardrails implementados (si GenAI) — input + output + operational
• CI/CD con validation gates configurados y testeados
• Blue & Gold deployment operativo con rollback verificado
• Monitoring stack completo: infrastructure, application, model, data
• Drift detection activo con alertas configuradas
• Runbooks documentados para failure scenarios comunes
• Cada fase tiene Definition of Done cumplida con evidencia

Cross-References

Output Format Protocol

if FORMATO == "ejecutivo":
  Stack summary + implementation timeline + resource needs + milestones
  Audiencia: Project managers, sponsors

if FORMATO == "técnico":
  Full 6-section implementation guide + blueprints + configurations
  Audiencia: ML engineers, DevOps, platform engineers

if FORMATO == "híbrido":
  Executive timeline + technical deep-dive completo
  Audiencia: Tech leads planning implementation sprints

Output Artifact

## {System Name} — AI Architecture Implementation Guide

### Stack Decision
[S1: Technology selection matrix with ADRs]

### Phase 0: Foundation
[Repository structure, CI skeleton, development environment]

### Phase 1: Data Pipeline
[S2: Pipeline implementation, quality gates, feature store]

### Phase 2: Model Development
[S3: Training pipeline, experiment tracking, model registry]

### Phase 3: Serving & Inference
[S4: Model serving, API, caching, guardrails, RAG/agents]

### Phase 4: CI/CD & Deployment
[S5: Blue & Gold pipeline, gates, canary, rollback]

### Phase 5: Monitoring & Operations
[S6: Observability stack, dashboards, alerts, runbooks]

### Implementation Timeline
[Gantt chart with milestones and dependencies]