Skip to content

Latest commit

 

History

History
503 lines (361 loc) · 15.6 KB

File metadata and controls

503 lines (361 loc) · 15.6 KB
marp true
theme default
paginate true
size 58140
header The AI Stack — from infrastructure to application
style section { font-size: 26px; } h1 { font-size: 44px; } h2 { font-size: 34px; } code { font-size: 0.8em; } pre { font-size: 0.7em; line-height: 1.25; } table { font-size: 0.78em; } section.lead h1 { font-size: 52px; } section.lead { text-align: center; } footer, header { color: #888; font-size: 14px; }

The AI Stack

From infrastructure to application


Agenda

  1. Why a stack? — a model is one piece
  2. The 5 layers — infra → models → data → orchestration → application
  3. The business view — revenue pyramid · 3 types of innovation
  4. Case studies — chatbot · RAG · the vendor-vs-operator trap
  5. GenAI for DevOps — incident RCA · observability · IaC · CI/CD risk
  6. Takeaways

Why a "stack"?

A model is one piece, not the whole system.

To solve a real problem you also need: compute to run it, data to ground it, orchestration to coordinate it, and an application for the user.

Every choice across the stack drives quality · speed · cost · safety.


The 5 layers

  5 · APPLICATION     interfaces & integrations — the user
  4 · ORCHESTRATION   plan → execute → review — the "brain" (agentic)
  3 · DATA            sources, pipelines, vector store, RAG
  2 · MODELS          open/proprietary · size · specialization
  1 · INFRASTRUCTURE  GPUs — on-premise, cloud, local

We'll go bottom → top, then look at the business picture.


1 · Infrastructure

LLMs need AI-specific hardware (GPUs).

Three ways to deploy:

  • On-premise — own it; full control, high upfront cost
  • Cloud — rent it; scale up/down on demand
  • Local — laptop; only the smaller models

The hardware you can access decides what models you can run at all.


2 · Models

Pick along three dimensions:

  • Open vs proprietary — control, cost, where it runs
  • Size — large (LLM) vs small (SLM)
  • Specialization — reasoning, tool-calling, code, language

Use vs build: consume a model, fine-tune it, or (rarely) train it. Lifecycle glue = MLOps.

Need fresh knowledge, not new behavior? Prefer RAG over fine-tuning.


3 · Data

Base models have a knowledge cutoff and don't know your private data.

sources → pipelines (clean/chunk) → embed → vector store → RAG → model

RAG retrieves relevant context and augments the prompt — fresh & private knowledge without retraining.


4 · Orchestration — the agentic layer

One prompt in / one answer out = a chat box. Add the loop and it becomes an agent:

plan ──▶ execute (tools) ──▶ review ──▶ (loop)
   ▲                                      │
   └──────────── memory / state ◀─────────┘

Two kinds: static (reliable workflows) + agentic (autonomous decisions). Fastest-moving layer — agents, MCP.


5 · Application

Where AI meets the user. Two questions:

  • Interfaces — text / image / audio · revisions · citations · conversational UI replacing forms
  • Integrations — inbound (tools feed AI) & outbound (AI output → tools)

"What the AI does" is orchestration. "What the user sees / where the result lands" is the application layer.


The business view

Where is the money — and where is the value?


Revenue pyramid (mid-2026)

              ┌──────────────┐
              │     APPS     │  fragmented startups
            ┌─┴──────────────┴─┐
            │  ORCHESTRATION   │  LangChain ~$16M ARR
          ┌─┴──────────────────┴─┐
          │        DATA          │  Scale AI ~$2B
        ┌─┴──────────────────────┴─┐
        │   FOUNDATION MODELS      │  OpenAI ~$33B · Anthropic ~$45B*
      ┌─┴──────────────────────────┴─┐
      │   INFRASTRUCTURE / CHIPS      │  Nvidia ~$75B/qtr → ~$300B run-rate
      └──────────────────────────────┘

Revenue concentrates at the base (Nvidia alone > all model companies). Hype ≠ revenue — orchestration gets the most attention, the least money (LangChain ≈ 1/18,000 of Nvidia).


Three types of innovation

Clayton Christensen — a lens for any AI product

Type Idea Jobs AI example
Market-creating expensive/hard → cheap for all creates foundation models for everyone
Sustaining make a good product better neutral an AI feature in your app
Efficiency same work, less cost reduces coding assistants, automation

The type isn't the layer — it's how the tech is used.


Case · sales / support chatbot

Klarna × OpenAI: 2.3M chats in month 1 = 700 agents' work, 67% automated, 11 min → <2 min, ~$40M profit (2024). Caveat: cut too deep, rehired in 2025 — AI-first, not AI-only.

ILLUSTRATIVE — 100,000 contacts/mo · human $5 · AI $1 · 67% deflection
  AI handles  67,000 × $1 = $67k/mo
  vs humans   67,000 × $5 = $335k/mo
  net saving  ≈ $268k/mo  ≈ $3.2M/yr

Players: Sierra · Decagon · Ada · Intercom Fin


Case · RAG over internal docs

Morgan Stanley × OpenAI

  • Indexed 350,000 documents (40M words) with RAG
  • Before: 30+ min manual search · advisors reached ~20% of knowledge
  • After: instant · 98% of teams use it · access 20% → 80%

Frees advisor time for client work; answers grounded in verified sources.

Similar: Glean · Harvey · Hebbia


The trap: vendor value vs operator value

The pyramid measures who earns money selling AI → base wins.

Case studies measure value from applying AI → lands on the operator's books.

Klarna's $40M isn't an "AI app company" revenue line — it's on Klarna's P&L. The app layer looks thin for vendors, yet is where operators win.


GenAI for DevOps


An AI agent for ops = the whole stack

  5 · APPLICATION    Slack on-call copilot · NL query · Datadog/PagerDuty
  4 · ORCHESTRATION  agent: query logs/metrics/traces → traverse deps → RCA
  3 · DATA           vector DB (FAISS/Weaviate): telemetry + incidents + runbooks
  2 · MODELS         LLM, prompt-engineered/fine-tuned for logs & remediation
  1 · INFRASTRUCTURE GPUs to run it

This is why we learned the stack: a production SRE assistant exercises all five layers at once.


Incident triage, RCA & remediation

An AI agent watches alerts, reads logs / metrics / traces, and proposes a root cause and a fix.

Tools: Cleric · Resolve.ai · Traversal · Rootly · incident.io · Datadog Bits AI

Proven in production:

Where Result
Traversal @ American Express 82% RCA accuracy · −32% MTTR
Microsoft RCACopilot ~0.77 RCA accuracy · 4 yrs, 30+ teams

Vendor "38–90% MTTR" claims are self-reported. Remediation stays human-in-the-loop.


Observability: ask telemetry in English

NL querying + anomaly detection on top of your existing tools (Datadog, Prometheus, Grafana, OpenTelemetry):

  • Datadog Bits Assistant — query dashboards/logs/traces in plain language
  • Grafana Assistant — NL telemetry questions + ML correlation
  • Datadog Toto — timeseries foundation model powering anomaly forecasting

Grounding = RAG over telemetry + incident history + runbooks in a vector DB (FAISS / Weaviate), so the LLM reasons over your system, not generic text.


IaC with AI — the real risk

The problem isn't bad code. It's plausible code — code nobody reads line-by-line anymore.

A 200-line Terraform module, AI-written in 30 s. The dev skims it, sees no syntax errors, runs apply. What goes unchecked:

  • Security group with ingress 0.0.0.0/0?
  • S3 bucket missing a public-access block?
  • RDS without encryption-at-rest? · IAM policy with Resource: "*"?
  • Secrets hardcoded in tfvars?

Anti-pattern: AI generates Terraform → apply → 💥 (no review, no scan, no policy gate)


What AI tends to get wrong

Common misconfigurations in AI-generated Terraform (approx.):

# Misconfiguration Seen in Why
1 Security group 0.0.0.0/0 on 22/3389 ~25% common demo examples
2 S3 missing public-access block ~20% not blocked by default
3 RDS not encrypted ~18% encryption is opt-in
4 IAM Resource: "*" ~15% convenient in demos
5 Secrets in tfvars (no SOPS) ~12% pattern is in training data
6 Missing tags ~30% tag policy is org-specific
7 Hardcoded creds in provider block ~5% rarer, but critical

The 3-layer defense

1 · AI GENERATE    Claude Code + Terraform MCP
                   constraint-first, security-first prompts
        ↓
2 · HUMAN REVIEW   read the plan, understand every resource
                   ask the agent what you don't get — never approve blindly
        ↓
3 · POLICY GATE    tflint → checkov → terraform plan → conftest
                   → AI explains the diff → human approves → apply
                   RULE: checkov not clean (any HIGH) → no apply

L1 is fast but plausible-but-wrong · L2 catches intent, but humans tire · L3 is automated & consistent — it catches what humans miss.


CI/CD risk & guardrails (the frontier)

Emerging, less mature than incident response:

  • Blast-radius analysis — LLM predicts what a change can break
  • Risk scoring of config / schema changes before rollout
  • Auto validation & rollback informed by historical outcomes

Honest take: fewer proven products here than in RCA/observability — strong territory for a platform team to build real differentiation.


Takeaways

  1. AI is a stack, not a model — infra → models → data → orchestration → app
  2. Money sits at the base; value from applying AI sits with the operator
  3. Name the innovation type: market-creating / sustaining / efficiency
  4. Always do the math with both sides — savings and the AI's own cost
  5. In DevOps, AI drafts — your pipeline verifies

Thank you

Questions?

Speaker notes & full numbers: README.md and docs/devops-cases.md