AI development glossary
Plain-English definitions for the terms we use across every build. Take any of them into your own team. Everything below is grouped by category, with cross-references where one term builds on another.
AI concepts
6 terms
RAG (Retrieval-Augmented Generation)
Pull facts from your data, then ask the model.
RAG is the pattern of retrieving relevant documents from your own data store before sending a prompt to the model. The retrieval step grounds the answer in your facts, which reduces hallucination and lets the model speak about content that was never in its training set. RAG is the right answer for most product questions about owning data; fine-tuning is the right answer when the failure mode is style or formatting.
Agent workflow
Multi-step LLM pipelines with retries and tools.
An agent workflow is a sequence of LLM calls where each call decides the next action: retrieve, summarise, draft, validate, retry. Production agent workflows have observability, retries, fallbacks, and a hard stop. The opposite of an agent is a single prompt that returns a single response. Agents earn their place when the task has multiple branches that depend on intermediate results.
Kill-switch
Disable an AI feature in production with one toggle.
A kill-switch is a feature flag wired specifically around an AI capability so it can be turned off without redeploying. Production AI features ship with a kill-switch because they fail differently than typical software: a model swap, a provider outage, or a regression in a prompt can degrade quality silently. The kill-switch lets you fall back to a non-AI baseline (cached answer, human handoff, classic search) while you investigate.
See also
Vendor parity
Every AI feature is swappable to another provider in a week.
Vendor parity means a feature is not tied to a single provider's idiosyncrasies. The prompt is structured so it can be re-tested against another provider; the eval suite is portable; the SDK calls go through an adapter. The cost of vendor parity is small (a thin abstraction); the cost of not having it is large the day a provider raises prices, deprecates a model, or hits an outage.
Embeddings
Numeric representations of text used for semantic search.
An embedding is a vector of floating-point numbers that represents the meaning of a piece of text. Two pieces of text with similar meaning land near each other in vector space. Embeddings power retrieval (the first step of RAG), clustering (group related items), and recommendations (find similar). Pick the embedding model first; the rest of the retrieval pipeline depends on it.
Eval suite
Repeatable tests that score a prompt against expected outputs.
An eval suite is the AI equivalent of a unit test. You define inputs and expected outputs (or expected properties of the output), and the suite scores prompts deterministically. Eval suites let you compare two prompts, two models, or two versions of the same prompt without subjective review. We require an eval suite on any AI feature that ships in a critical path.
Engineering practice
3 terms
Code review prompt
AI first-pass review applied to every PR before a human looks.
A code review prompt is a structured template that runs against every pull request. It catches obvious bugs, missing edge cases, security smells, and dead code before a senior engineer's time is spent on the review. It does not replace senior review; it makes senior review faster.
See also
Managed infrastructure
Vendor-hosted services with zero-ops as the default.
Managed infrastructure means renting the runtime: Vercel for hosting, managed Postgres for data, Resend for email, Vercel AI for inference proxying. The opposite is self-hosted: your own Kubernetes, your own database tuning, your own paging rotation. Managed beats self-hosted until you have a real reason to leave; the exit cost compounds.
See also
Boilerplate
Repetitive scaffolding that AI generates well.
Boilerplate is the routine scaffolding of a build: routes, types, forms, schemas, CRUD handlers, basic test stubs. AI is excellent at boilerplate, which is one of the largest single sources of speed-up. Boilerplate is the worst place to spend senior time, and the best place to apply automation.
Engagement model
7 terms
Sprint-fixed pricing
One price per sprint, written scope, no hourly billing.
Sprint-fixed pricing is a contract where each sprint has a defined scope, a fixed price, and a fixed timeline. The agency, not the client, carries the schedule risk. Scope creep caused by the agency is on the agency. Hourly billing inverts this: every estimate becomes a baseline and the schedule risk transfers to the client.
See also
Mutual NDA
NDA that protects both sides equally.
A mutual NDA is a non-disclosure agreement that obliges both parties (client and agency) to the same confidentiality terms. Most off-the-shelf NDAs only protect one side, the side that drafted it. We sign a mutual NDA before reading a word of your product, before the discovery workshop, and before the audit.
Discovery workshop
Paid, refundable scoping session before any sprint.
A discovery workshop is a 2 to 4 session paid engagement that turns a product idea into a signed scope, a cost model, and a go or no-go recommendation. It is refundable: if the result is no-go, the deposit is returned. Discovery is what separates a serious agency from one that quotes from the homepage form.
Product audit
Free 48-hour review by a senior engineer.
A product audit is a written PDF and a Loom from a senior engineer, delivered within 48 business hours of submission. It identifies three concrete AI integrations with cost and payback, one "don't build this" recommendation, and where the product is healthy versus brittle. The audit is free, capped, and not a sales call.
Senior-only delivery
Every engineer on the project has 5+ years of production experience.
Senior-only means no juniors, no contractors, no offshore handoffs. Each engineer who touches the codebase has at least five years of shipping production software. The cost premium is real; the time-to-resolve, the design quality, and the long-term maintainability gain offsets it within the first sprint.
See also
Post-launch guarantee
30 days where bugs we shipped are bugs we fix, no clock.
After launch, anything we shipped that is broken gets fixed inside 30 days at no charge. No surprise invoice, no clock. New features after launch are a retainer, not a guarantee.
See also
SOC-2-friendly delivery
Practices that align with SOC-2 controls without the certification.
SOC-2-friendly means the engagement runs with audit logs, access controls, mutual NDA, defined data residency, and incident reporting. We are not a SOC-2-certified body, but the engagement does not require you to weaken any of your existing controls to work with us.
See also
Build process
4 terms
Methodology library
Public archive of prompt templates with model and time saved.
Our methodology library is the public archive of every prompt template we use across builds. Each template names the model, the phase of the build it applies to, the human time it replaces, and the changelog. The library is updated monthly as models evolve.
AI touchpoint
A specific point in the build flow where AI is applied with a human owner.
An AI touchpoint is a documented point in the process where AI assists a human owner: discovery synthesis, scope drafting, code review, retro synthesis, doc first-pass. Every touchpoint has a name, a model, an input shape, and a human who is responsible for the output.
Weekly demo
A working demo every Friday: Loom, metrics, live URL.
Every sprint week ends with a working demo: a Loom walkthrough, the metrics dashboard, and the live URL. No slide decks, no redacted screenshots. If a Friday demo is missed, the week is on the agency.
See also
Speed Report
Quarterly publication on time-to-ship, with raw data.
The Speed Report is our public audit of how fast we ship, normalised against scope. Each quarter we publish the matched-cohort comparison, the per-phase breakdown, what we measured, what we excluded, and where the data is honestly thin.
See also
Free, 48-hour SLA, no sales call
Apply these terms to your specific product
The free Product Audit returns where each of these concepts applies to your build, what is wired correctly, and what is missing. Senior engineer, 48-hour turnaround.