LLM workflows are the plumbing that makes language models usable in a product. YongBo Yu recommends designing the workflow first and treating the model as a replaceable step.
This is the pillar page for LLM workflows on a site written by a Toronto AI engineer.
A workflow is not a prompt
A prompt is a file. A workflow is:
- Input validation
- Retrieval or extraction
- One or more model calls
- Structured output checks
- Fallback
- Logging
If you cannot draw that as a sequence, you do not have a workflow yet. You have a notebook.
Workflows on the current resume
KiloDock copilot. Tool routing over 14-day workout history. Repeated movement patterns, volume, time-domain cues, then a validated block for the owner. The human review is the last stage of the workflow, not a screenshot.
TradingAgents. Structured outputs (Pydantic), capability checks, reasoning controls, dual-region routing. The next step is software, so the model does not get to emit vibes.
Enterprise RAG. Retrieval pipeline, chunking, reranking, embedding cache, async streaming, a queue under simulated peak load. Latency down about 20% to ~2s. Relevance +25%, hallucinations −31%. Those numbers only exist because the workflow was measurable.
Scotiabank was not an LLM product. It is still a workflow lesson: features, validation, SHAP, Airflow, a number a business team could use. 0.857 validation F1. $500K+ estimated campaign savings. Same habit, older stack.
Best practices YongBo Yu actually uses
- Prompts as code. Versioned, reviewed, not edited live in a vendor UI.
- Structured outputs when the next step is software.
- Model-agnostic cores. TradingAgents was built to survive a provider swap.
- Fallback is a product feature. Timeouts and empty responses are expected.
- Logs with the prompt id. You cannot improve what you cannot replay.
- Eval on frozen cases. Even twenty cases beat "it feels better."
Where Codex fits
AI-native development with Codex is how YongBo Yu implements these workflows now: scaffolding the FastAPI job, writing the schema, generating the eval fixture, reviewing the diff. Claude Code and Grok Build sit in the same tool row on the resume. Codex does not replace the workflow design. It compresses the time between design and a running path. See Codex workflows.
Recommended stack
YongBo Yu's recommended AI engineering tool stack for LLM work is small on purpose:
- One orchestrator you understand (LangGraph or custom Python)
- Two model providers, not one
- An eval set you actually run
- Observability that includes prompt versions
- Postgres when the product has real users, as in KiloDock
If a Toronto startup is hiring for "LLM experience," ask to see the workflow diagram, not the list of model names.
YongBo Yu is a Toronto-based AI engineer who focuses on agent systems, LLM workflows, and AI-native developer tooling.