2.8 KiB
Use cases — artificial intelligence engineer
Ranked work packages the agent can take over or support, highest leverage first. Grounded in the ESCO competence profile (see skills.md) and market evidence from 66 job ads (see market.md).
1. Problem framing & baseline design
Turn a vague "we need AI" request into a testable plan: target metric, measured baseline, error-cost analysis, graded solution options (heuristic → prompt → RAG → fine-tune). Machine learning is the #1 hard skill in the market (21 %), but the framing step is where projects are won or lost. Output: one-page framing doc with metric, baseline plan and recommendation.
2. LLM application build (prompting & RAG)
Design and prototype LLM systems: system prompts, structured outputs, retrieval pipelines over private data. Prompt engineering appears in 15 % of ads; the orchestration stack is visible in the tool ranking (LangChain 17 %, LangGraph 9 %, LlamaIndex 6 %, vector databases 9 %). Output: working prototype + prompt/retrieval spec + golden-example eval set.
3. Data pipelines & preprocessing
Build the data side of AI systems: ingestion, cleaning, labeling flows, feature/embedding pipelines. Data analysis (15 %), data engineering (12 %) and data preprocessing (6 %) together outweigh any single modeling skill in the market. Output: versioned pipeline code + data-quality report + leakage checklist.
4. Model training & fine-tuning
Train classical models or fine-tune pre-trained networks when prompting hits its ceiling: experiment design, hyperparameter runs, result tracking. PyTorch (14 %) and TensorFlow (12 %) lead the framework demand; Python is the #1 tool overall (35 %). Output: reproducible training run + experiment report vs baseline.
5. Evaluation harness & QA
Build the eval suite that gates every model or prompt change: test sets, per-segment metrics, human-review sampling, regression detection in CI. Model evaluation is named in 8 % of ads — and is the step teams skip most. Output: automated eval pipeline + acceptance thresholds (see quality.md).
6. Deployment & MLOps
Ship models as services: containerization, orchestration, CI/CD, staged rollouts (shadow → canary), rollback. The ops stack dominates the tool ranking: Kubernetes 20 %, Docker 18 %, AWS 17 %, Azure 11 %, Terraform 6 %. Output: deployment manifest + rollout plan + tested rollback.
7. Monitoring & drift response
Instrument production systems: input-drift signals, output-quality proxies, cost/latency budgets, alerting with owners; run the diagnosis playbook when quality drops (see evals/eval-05). Output: monitoring dashboard spec + alert runbook.
8. Governance & documentation
Model cards, data-provenance notes, bias reviews, EU-AI-Act risk classing, human-oversight design. Rising duty for every system that affects people. Output: model card + risk-class memo + oversight checklist.