“AI/ML可观测性”

ML Deployment / Serving Architect (Online & Batch)

Serving design: real-time vs batch vs streaming, latency budget, feature parity, shadow rollout, and resource sizing.

Model Monitoring / Drift Detection Designer

Production ML monitoring: data/concept drift, staleness, fairness over time, and retraining triggers — with alert thresholds tied to impact.

Exploratory Data Analysis (EDA) Conductor

Structured EDA with statistical thinking — distributions, leakage, correlations, quality issues — ends with modeling implications, not pretty-but-empty charts.

Feature Engineering Transformer Designer

Production-safe feature pipelines: fitting on train only, consistent train/serve, and leakage-free transforms.

Model Validation / Experiment Designer (CV Strategy)

Chooses correct validation (grouped/time-series/nested CV), prevents over-opt, and reports uncertainty — the difference between local scores and prod results.

Model Training / Hyperparameter Optimization Engineer

Rigorous training loops: reproducibility, resource accounting, sane HPO, and checkpointing — for classical and deep models.

Model Evaluation, Fairness & Calibration Auditor

Beyond accuracy: calibration, subgroup performance, error types, and decision thresholds — what compliance/trust reviews actually need.

Data Pipeline / ETL Engineer (Spark/dbt/SQL)

Reliable analytics engineering: layered models, tests, freshness, and idempotent loads in dbt/Spark.

Data Warehouse / Lakehouse Modeler

Warehouse schema design for scale and analytics: clustering, partitioning, semantic layer, and cost governance.

LLM Fine-Tuning / Model Customization Planner

Decides between prompting, few-shot, RAG, and fine-tuning; if fine-tuning, designs dataset, method (LoRA), and eval.

Embedding / Vector Search Engineer

Chooses embeddings and ANN indexes correctly: dimensions, chunking, hybrid search, and recall tuning.

Prompt/Template CI & Regression Harness Builder

Productionizes prompt changes: versioned prompts, offline eval gates, online shadows, and rollback — prompt-as-code.

Synthetic Data Generator / Augmentation Designer

Privacy-preserving synthetic data for testing/training with fidelity and utility checks (avoid fake data that leaks or misleads).

A/B Experimentation Platform & Statistics Advisor

Correct experiment design: power, guardrails, SRM, peeking, and analysis of metric movements.

Time-Series / Forecasting System Engineer

Builds forecasting pipelines with correct validation, uncertainty, and operational regime handling (promotions/outages), not just model fitting.

Data Governance / Privacy Engineering (GDPR)

Implements privacy by design: inventory, minimization, consent, DSAR pipeline, and deletion across systems.

Data mapping: inventory of personal data per system (collection points, purpose,

Minimization & controls: collect least data, pseudonymize/hash identifiers in

User rights pipeline (DSAR): unified identity resolution across services;

Consent & cookies: consent management (strictly-necessary separation), granular