Fieldwork, value, engineering, and operations for AI that works beyond the demo.
An open-source design and verification kit for applied-AI engineers, product and workflow owners, service operators, and forward-deployed engineers (FDEs).
Five-minute field guide · Complete method · One worked engagement · 12 Factors of AI Value Engineering · Executive funding guide
| You need to… | Start here | Leave with |
|---|---|---|
| Decide where AI could help | Discovery and Value, then mechanism selection | An observed decision, accepted outcome, verifier, value hypothesis, and simplest plausible mechanism |
| Find a first workflow | First workflow investigation and the discovery pack | Recent cases, system path, and unknowns—not a quality claim |
| Build a useful first feature | Build one vertical slice and try the invoice review lab | One bounded path through real inputs, a usable review surface, failure behavior, and test cases |
| Fix weak retrieval or unsupported answers | Context and knowledge systems and the retrieval evaluation lab | Separate evidence for retrieval, source support, permissions, freshness, abstention, latency, and cost |
| Make a retry or recovery path safe | Enterprise integration reality and the durable recovery lab | A stable operation identity, explicit ambiguity, source-of-truth readback, and an escalation path instead of a duplicated effect |
| Test a model, prompt, source, or policy change | Wrong-amount change walkthrough, then production evaluation and change impact | A source-owned case, candidate fix, neighboring regressions, dependencies, and a release decision |
| Launch or operate a system | Production readiness and production operations | The missing evidence or capability, its owner, and a release, repair, constraint, pause, or retirement decision |
| Rescue a brief that does not match the work | Five-minute Guide and field engagement and reframing | One representative case, conflicting claims, a safe fallback, and the person who may decide |
| Learn by doing | Invoice practice packet | A source-linked decision, a deliberate failure, a review result, and the limits of the evidence |
| Layer | Use it for | Entry |
|---|---|---|
| The Guide | The mental model and canonical delivery loop | Five-minute Guide, then concise Guide |
| Handbook | Running a live engagement | Lifecycle playbooks |
| Engineering Kit | Contracts, controls, architecture, evaluations, operations, and executable evidence | Templates, controls, and examples |
They are not separate frameworks. Use the capability roadmap to plan practice.
Start with the work and the accepted outcome, not a model or agent topology. Compare deterministic software, optimization, classical ML, retrieval, a foundation-model call, a bounded agent workflow, and human review. Choose the smallest mechanism that can safely do the job.
Tokens are an input. Autonomy is a design choice. Accepted outcomes are the product.
Use the 12 Factors and one-page scorecard to test the outcome, verifier, adoption, authority, cost, and proof. The executive funding route supports the investment decision.
Try the invoice practice packet: a messy brief, difficult documents and review rubric. Its runnable lab compares rules with optional model proposals.
The invoice-exception engagement follows a sold promise that field evidence kills: reframe, economics, controlled-write runtime, retrieval comparison, restart-safe recovery practice, evaluation, blocked handoff, and review-only decision.
The shipment-risk example combines classical ML, deterministic routing, optional model explanation, and human review.
The finance variance-commentary walkthrough shows a review-first path: code calculates, owners explain, finance approves the model draft.
The service-referral case separates a reported business experiment from fictional practice with blocked work and delayed payments.
npm ci --ignore-scripts
npm run test:retrieval-evaluation
npm run test:durable-recovery
npm run test:reference
npm run test:evals
npm run test:hybridThe runnable examples use simulated services; the durable-recovery lab also persists local state. Passing tests proves only the declared local behavior—not customer value, production readiness, or deployment approval. The finance and referral walkthroughs are reading exercises.
Before adapting them, use Enterprise Integration and Scale Reality to replace teaching conveniences with target evidence.
The canonical lifecycle:
flowchart LR
A["Understand the request and workflow"] --> B["Observe and reconcile the work"]
B --> C["Charter value and scope"]
C --> D["Make data fit for the decision"]
D --> E["Select the mechanism"]
E --> F["Build one controlled slice"]
F --> G["Prove it with cases and users"]
G --> H["Launch with operating ownership"]
H --> I["Operate, learn, or retire"]
Each transition needs evidence and an accountable decision. A model score, sponsor, or deadline cannot override a failed value, authority, safety, ownership, or production gate.
Validate a working artifact before it is complete:
npm run validate:artifact -- ./path/to/workflow-start.json --profile starter --type workflow-charter
npm run validate:artifact -- ./path/to/workflow-charter.json --profile completeThe starter profile checks the few fields needed for the current decision while retaining the same canonical types and closed-object rules. It is not a second schema. See artifact validation.
After the workflow and value are accepted, choose a business-flow pattern and, when useful, an industry profile. The solution portfolio is a design hypothesis, not target evidence or a deployable product.
The guide is complete as documentation. Sixteen optional skills provide focused routes over the same canonical artifacts:
npx skills add davidahmann/applied-ai-field-guidePin the source. Skills grant no authority or evidence. Give an agent AGENTS.md.
The Applied AI Field Guide local plugin adds local continuity for sources, revisions, decisions, dependencies, and review packets. Keep restricted content in its approved source system; local execution is not permission to copy it.
Describe the situation in ordinary language and confirm which skill the host selects.
The control catalog is project policy, not an external compliance standard. Target organizations retain architecture, risk, and release authority.
Contributions should improve an existing route before adding another one. See CONTRIBUTING.md, repository maintenance, security policy, the Apache-2.0 license, and third-party notices.
Created and maintained by David Ahmann, a cloud, data, and applied-AI platform leader with Field CTO experience. I’m building the guide to make the decisions behind useful AI systems easier to learn, test, and put into practice. It combines delivery experience, source-linked research, and clearly labeled teaching examples; it is independent work, not employer guidance or endorsement.