Skip to content
awesome-applied-ai

A taxonomy of applied AI

The applied AI cycle, filed bystage, layer, problem, equation, tool

What an enterprise system actually goes through between someone asking for it and someone signing off on it, in 8 stages. Under each stage: the design problems it raises, the arithmetic that settles them, and the tools worth knowing.

Try
cycle stages
8
cycle stages
design problems
20
design problems
layers
6
layers
equations
20
equations
index entries
133
index entries
defined terms
121
defined terms
commands
126
commands
kit entries
73
kit entries
build projects
8
build projects
case studies
5
case studies

The author · July 2026

Built by an Applied AI Specialist shipping agent systems in production

Vahid Faraji is a Senior Applied AI Specialist at Kariyer.net (ilab group). The map above comes from the same notes as his context engineering toolkit and the multi-agent systems he runs at enterprise scale.

Visit the portfolio →LinkedIn

The cycle

all 8 stages →
  1. 01FrameWhat decision changes when this ships?no single layer · 1 problem
  2. 02SourceWhere does the data actually live, and what may you do with it?retrieval · 2 problems
  3. 03GroundHow does a query become evidence the model can use?retrieval · 6 problems
  4. 04AssembleWhat goes in the window, in what order, at what price?caching · 2 problems
  5. 05OrchestrateOne call, a chain, or several agents?orchestration · 6 problems
  6. 06ProveIs the new thing better than the old thing on your data?evaluation · 2 problems
  7. 07ServeWhat does it cost at real concurrency, and how fast is P95?caching · 1 problem
  8. 08GovernWho signs off, what is logged, and how do you replay a bad answer?governance · 2 problems

Start here if

You are shipping this at work

Walk the 8 stages in order. Each one names the number you have to compute, the failure that shows up if you skip it, and the problems filed under it.

The cycle →

Start here if

You are building toward the job

8 projects, each small enough for a weekend and each ending in a number you measured yourself. They map onto the same 8 stages.

The build track →

What settles the argument

all 20 equations →

Retrieval

HNSW resident memory

M  ≈  N (4d+8m)M \;\approx\; N\,(4d + 8m)

Ten million 1536-dimension vectors at m = 16 want roughly 63 GB resident. That figure picks your architecture. Query latency does not.

Caching & Context Optimization

Effective input cost at hit rate h

Ceff  =  Cbase[(1−h) w+h r]C_{\text{eff}} \;=\; C_{\text{base}}\left[(1-h)\,w + h\,r\right]

At a 90 percent hit rate on a five minute cache, input spend falls 78.5 percent. Hit rate is a larger lever than the model you chose.

Orchestration & Protocols

Reliability compounds multiplicatively

Psuccess  =  ∏i=1npi  =  p nP_{\text{success}} \;=\; \prod_{i=1}^{n} p_i \;=\; p^{\,n}

Twenty steps at 99 percent each succeed 81.8 percent of the time. At 95 percent each, 35.8 percent. No amount of prompting closes that.

Measured in production

5 case studies →

The stack under it

why six →
01Retrieval4602Memory & State1303Caching & Context Optimization1504Orchestration & Protocols2805Evaluation & Observability1606Guardrails & Governance15

Everything else

Problems20Dictionary121Commands126Research6Shipped5Instruments3Skills1Kit73Index133

Every layer page carries its own arithmetic next to my notes on where the received wisdom is wrong: stalled projects still appearing in roundups, standards that are not standards yet, controls that are theatre. Read those before the tool lists. Or search all 133 entries across 147 topics.