Data strategy & engineering for agentic workflows

by Ray Grieselhuber

Embed

Start on current slide

Slide 1

Slide 1 text

Data strategy & engineering for agentic workﬂows Ray Grieselhuber Founder & CEO, DemandSphere @MKE DMC April 15, 2026

Slide 2

Slide 2 text

Agenda ● Intro ● Background and motivations for this topic ● Vibe coding vs. AI-driven engineering ● Tools & skills ● Testing & security ● Project ideas ● Open weight models vs. closed models ● Final thoughts

Slide 3

Slide 3 text

Introduction ● BA, Japanese & B.S. Computer Science ● Software engineering with focus on early ML, agents, big data, distributed systems, etc. ● SEO since 2006, built ﬁrst enterprise platform on the market ● Founded DemandSphere (GinzaMetrics) in 2009, Tokyo ● Moved to Silicon Valley in 2010 for YC ● Work with clients all over the world

Slide 4

Slide 4 text

Quick description of our platform for context

Slide 5

Slide 5 text

Today we’re building a map, don’t worry too much

Slide 6

Slide 6 text

Goal is to promote more ﬂuency in building with AI

Slide 7

Slide 7 text

Manager mode vs. Founder mode

Slide 8

Slide 8 text

We are all in Founder mode now

Slide 9

Slide 9 text

Founder mode means: ● Staying close to the user (even if it’s yourself at ﬁrst) ● Build what people want ● Iterate quickly ● Keep your sleeves rolled up

Slide 10

Slide 10 text

Software as a Service Or Service as a Software?

Slide 11

Slide 11 text

Experience at Functionize taught me about the new SaaS

Slide 12

Slide 12 text

The new SaaS (Service as a Software): ● AI is empowered by human attention ● Businesses pay for results, they don’t care about your cool AI systems if they don’t work ● Humans have to backstop when AI systems fail

Slide 13

Slide 13 text

The dirty little secret in SaaS

Slide 14

Slide 14 text

Silicon Valley is ﬁnally starting to ﬁgure this out

Slide 15

Slide 15 text

Vibe coding vs. AI-driven engineering

Slide 16

Slide 16 text

Don’t look down on vibe coding

Slide 17

Slide 17 text

Vibe coding is good for building prototypes and testing ideas quickly

Slide 18

Slide 18 text

AI-driven engineering is for building products

Slide 19

Slide 19 text

Vibe coding should be the starting point

Slide 20

Slide 20 text

Deployments and operations are what separate vibe coding and AI-driven engineering

Slide 21

Slide 21 text

People and processes (manual & automated) support actual products

Slide 22

Slide 22 text

Simple example: AI frontier model tracker

Slide 23

Slide 23 text

Constant updates + QA (human & machine)

Slide 24

Slide 24 text

Cambrian explosion of tools and ideas

Slide 25

Slide 25 text

Need to anchor to business goals

Slide 26

Slide 26 text

But the cost of experimentation is cheaper

Slide 27

Slide 27 text

Let’s talk about some basics

Slide 28

Slide 28 text

Tokens are the currency of AI and heavy context spends more tokens

Slide 29

Slide 29 text

Tokens = cost + latency + attention budget

Slide 30

Slide 30 text

The best AI development workﬂows: ● Compress ● Summarize ● Deduplicate ● Retrieve only what is needed

Slide 31

Slide 31 text

It’s about curation of the context window

Slide 32

Slide 32 text

(This also applies to AI search & SEO btw)

Slide 33

Slide 33 text

Tools you should use & know

Slide 34

Slide 34 text

Skills (also an open standard)

Slide 35

Slide 35 text

Be very careful with skills you did not build (supply chain attack)

Slide 36

Slide 36 text

Claude Code

Slide 37

Slide 37 text

Working in a team environment

Slide 38

Slide 38 text

Context ﬁles

Slide 39

Slide 39 text

CLAUDE.md

Slide 40

Slide 40 text

CLAUDE.md ﬁles can be hierarchical

Slide 41

Slide 41 text

Per branch

Slide 42

Slide 42 text

Per subfolder

Slide 43

Slide 43 text

Have Claude Code generate code maps for context

Slide 44

Slide 44 text

Have Claude Code generate ERD diagrams for databases

Slide 45

Slide 45 text

Mermaid and Mermaid.live are great for diagramming

Slide 46

Slide 46 text

The point is that maintaining context for humans is just as important as maintaining it for agents

Slide 47

Slide 47 text

Simple text “standards” are always the thing to bet on

Slide 48

Slide 48 text

HTML Vanilla JS JSON CSV Prompts Git (based on hashes) Markdown (.md) Mermaid

Slide 49

Slide 49 text

Monitor your /context and /compact when needed

Slide 50

Slide 50 text

A basic vibe coding to AI-driven engineering workﬂow Start in Claude.ai Vanilla prototype Claude.ai to build CLAUDE.md Move to local filesystem claude init Create repo Enforce testing policy Configure deployment environment Automate deployments Deploy regularly

Slide 51

Slide 51 text

Let’s talk about building MCP servers

Slide 52

Slide 52 text

MCP servers start oﬀ as local and “self-hosted” until you deploy them

Slide 53

Slide 53 text

Remember the diﬀerence between vibe coding and AI-driven engineering? Deployment!

Slide 54

Slide 54 text

Security & good tool construction are the hardest parts

Slide 55

Slide 55 text

Be sure to create an automated test suite

Slide 56

Slide 56 text

Security risks with MCPs: ● Client side ○ Giving too much access to the MCPs ○ Prompt injection ○ Data / key leakage ● Server side ○ Validating tool inputs ○ Sanitizing output ○ Rate limits ○ Access controls

Slide 57

Slide 57 text

Let’s get into solving some real problems

Slide 58

Slide 58 text

In search, we are trying to get into the context windows of our audience’s minds

Slide 59

Slide 59 text

Context windows are everywhere

Slide 60

Slide 60 text

This means we need lots of data about their behavior and motivations

Slide 61

Slide 61 text

GSC Analytics data Search volume SERP data Log ﬁles LLM data Query fanouts (the list is always growing)

Slide 62

Slide 62 text

The ﬁrst step is deciding where this data is going to live

Slide 63

Slide 63 text

Then how to get it

Slide 64

Slide 64 text

APIs by themselves have challenges

Slide 65

Slide 65 text

MCPs can help to make tool usage of APIs more eﬃcient

Slide 66

Slide 66 text

But MCPs are transient

Slide 67

Slide 67 text

API direct to dashboards introduces latency

Slide 68

Slide 68 text

API direct to agents is an improvement but granular

Slide 69

Slide 69 text

You can solve a lot of problems with data warehouses

Slide 70

Slide 70 text

BigQuery, Clickhouse, and Snowﬂake

Slide 71

Slide 71 text

BigQuery is probably the best default choice

Slide 72

Slide 72 text

Using BQ as a data buﬀer between systems

Slide 73

Slide 73 text

Knowledge bases for agentic workﬂows

Slide 74

Slide 74 text

Context ﬁles vs. LLM wikis vs. RAG

Slide 75

Slide 75 text

Entity canonicalization is a really hard problem

Slide 76

Slide 76 text

Final thoughts

Slide 77

Slide 77 text

AI feudalism is coming

Slide 78

Slide 78 text

Make the investment to own your own compute capabilities

Slide 79

Slide 79 text

Learn how to work with open weight models

Slide 80

Slide 80 text

Going to SEO Week? https://luma.com/p7r3nbvk

Slide 81

Slide 81 text

Join us at FOUND Tokyo 2027

Slide 82

Slide 82 text

Thank you!