Share
Embed
Slide 1
Slide 1 text
LLMs as domain experts
Slide 2
Slide 2 text
Talysson Oliveira
Software architect
& Chief Learning Officer
at Codeminer42
@talyssonoc
talysson.com
Slide 3
Slide 3 text
dddbrasil.com.br
linkedin.com/company/dddbrasil
youtube.com/@ddd-brasil
Slide 4
Slide 4 text
⚠
This is not a "how to drive a project with AI" talk
Also not an "AI is bad/good" talk
talysson.com
Slide 5
Slide 5 text
Can LLMs be used as domain experts?
talysson.com
Slide 6
Slide 6 text
Can LLMs be used as domain experts?
Can LLMs automate DDD practices?
talysson.com
Slide 7
Slide 7 text
Automating Domain-Driven Design:
Experience with a Prompting Framework
Eisenreich, T., Jusic, H., & Wagner, S. (2026, March 27) https://arxiv.org/abs/2603.26244v1
Slide 8
Slide 8 text
Requirements Input
Ubiquitous Language
Establishment
+
Event Storming
Simulation
System Prompt:
DDD Specialist
and
Architecture
Sparring Partner
Bounded Contexts
Identification
Aggregate Design
Technical Architecture
Mapping
Eisenreich, T., Jusic, H., & Wagner, S. (2026, March 27) https://arxiv.org/abs/2603.26244v1
Domain Model
Slide 9
Slide 9 text
Ubiquitous Language
Establishment
Event Storming
Simulation
Errors accumulate
without guidance
Bounded Contexts
Identification
Aggregate Design
Technical Architecture
Mapping
Eisenreich, T., Jusic, H., & Wagner, S. (2026, March 27) https://arxiv.org/abs/2603.26244v1
Slide 10
Slide 10 text
Ubiquitous Language
Establishment
Event Storming
Simulation
Errors accumulate
without guidance
Bounded Contexts
Identification
Aggregate Design
Technical Architecture
Mapping
Eisenreich, T., Jusic, H., & Wagner, S. (2026, March 27) https://arxiv.org/abs/2603.26244v1
Slide 11
Slide 11 text
Can LLMs automate DDD practices?
Some of them
talysson.com
Slide 12
Slide 12 text
Should we automate these DDD practices?
talysson.com
Slide 13
Slide 13 text
Ubiquitous Language
Establishment
Builds a shared understanding of the model
Event Storming
Simulation
Errors accumulate
without guidance
Bounded Contexts
Identification
Aggregate Design
Technical Architecture
Mapping
talysson.com
Requires understanding of the model
Slide 14
Slide 14 text
Ubiquitous Language
Establishment
If you fully automate these
Event Storming
Simulation
Errors accumulate
without guidance
Bounded Contexts
Identification
Aggregate Design
Technical Architecture
Mapping
talysson.com
You won't be able to guide these
Slide 15
Slide 15 text
Does it mean we should not use AI with DDD?
No, but humans should be active,
especially in the groundwork
talysson.com
Slide 16
Slide 16 text
What about SDD?
Spec-Driven Development
talysson.com
Slide 17
Slide 17 text
Idea
Specification
Implementation
Refinement
talysson.com
Validation
Slide 18
Slide 18 text
"The code is a lossy projection from the specification"
AI Engineer. (2025, July 12). The New Code - Sean Grove, OpenAI. https://www.youtube.com/watch?v=8rABwKRsec4
Slide 19
Slide 19 text
A set of properties a useful specification needs
-
Communicates meaning and intent
-
Uses a consistent and unambiguous vocabulary
-
Define a scope to which it applies
talysson.com
Slide 20
Slide 20 text
DDD spent 20 years solving a problem
SDD just discovered it has:
Effective and meaningful communication
talysson.com
Slide 21
Slide 21 text
We just need to write very
detailed specifications!
… right?
talysson.com
Slide 22
Slide 22 text
"[...] would be an utterly simple person […]
if he thinks written words are of any use except to
remind him who knows the matter about which they are
written, […] if you question them [the words], wishing to
know about their sayings, they always say only one and
the same thing [...]"
- Phaedrus, Plato, 370 BC
talysson.com
Slide 23
Slide 23 text
"[...] the programmer's knowledge transcends
that [knowledge] given in documentation [...]"
- Programming as Theory Building, Peter Naur, 1985
talysson.com
Slide 24
Slide 24 text
"The code is a lossy projection from the specification"
talysson.com
Slide 25
Slide 25 text
"The code is a lossy projection from the specification"
The specification is a lossy projection from the model
talysson.com
Slide 26
Slide 26 text
The model isn't a document or the code
It is shared understanding and intuition
talysson.com
Slide 27
Slide 27 text
Where is AI useful when practicing DDD?
talysson.com
Slide 28
Slide 28 text
LLMs as domain experts
talysson.com
Slide 29
Slide 29 text
LLMs as domain experts
Why would we want LLMs as domain experts?
talysson.com
Slide 30
Slide 30 text
To represent a problem + solution, we need a domain model
To explore a domain model, we need domain knowledge
To have domain knowledge, we need domain experts
And sometimes, domain experts may not be available
talysson.com
Slide 31
Slide 31 text
Why would domain experts not be available
-
Their available time is scarce
-
The access to them is not easy or practical
-
The communication is not in tune yet
-
The team doesn't have domain experts (yet)
talysson.com
Slide 32
Slide 32 text
How can LLMs help us in these cases?
talysson.com
Slide 33
Slide 33 text
Ubiquitous language
Domain experts
talysson.com
Development
experts
Slide 34
Slide 34 text
Domain experts
talysson.com
Development
experts
Slide 35
Slide 35 text
Domain experts
talysson.com
Development
experts
Slide 36
Slide 36 text
Domain experts
Development
experts
We do this first
talysson.com
Slide 37
Slide 37 text
To make the
best out of this
Domain experts
Development
experts
We do this first
talysson.com
Slide 38
Slide 38 text
Example:
Special in-flight meals
talysson.com
Slide 39
Slide 39 text
In-flight meals
talysson.com
Slide 40
Slide 40 text
talysson.com
Slide 41
Slide 41 text
talysson.com
Slide 42
Slide 42 text
You can't "vibe" your way to something you
can't recognize when you get there
talysson.com
Slide 43
Slide 43 text
What problem are we solving?
talysson.com
Slide 44
Slide 44 text
Special in-flight meals system
-
Allow airlines to offer special in-flights meals paid in advance
-
Customers order special meals up to 5 hours before the boarding
-
A list of meals is sent to a specialized company, which will cook and deliver
them directly to the airports
-
Customers can rate their experience and be part of a loyalty program and get
discounts and gifts
talysson.com
Slide 45
Slide 45 text
In-flight meals
Logistics
talysson.com
Slide 46
Slide 46 text
talysson.com
Slide 47
Slide 47 text
talysson.com
Slide 48
Slide 48 text
talysson.com
Slide 49
Slide 49 text
talysson.com
Slide 50
Slide 50 text
In-flight trolleys
(meal carts)
Turnaround
schedules
Extra in-flight
meals
Cold chain
Long-haul flights
talysson.com
HACCP
(Hazard Analysis Critical Control Point)
In-flight meal
shortage
Standardization
Slide 51
Slide 51 text
Then, ask about these concepts
in the same conversation
(within the same LLM context)
talysson.com
Slide 52
Slide 52 text
talysson.com
Slide 53
Slide 53 text
talysson.com
Slide 54
Slide 54 text
HACCP
Domain experts
talysson.com
HACCP
Development
experts
Slide 55
Slide 55 text
HACCP
Domain experts
talysson.com
Development
experts
Slide 56
Slide 56 text
Use LLMs to prime yourself for
conversations with domain experts
talysson.com
Slide 57
Slide 57 text
But there are two catches!
talysson.com
Slide 58
Slide 58 text
1
An LLM will give you the
industry-average knowledge
This knowledge/vocabulary might
not work for your case
talysson.com
Slide 59
Slide 59 text
2
LLMs are sycophants
"You're absolutely right!"
The "tone" you use with LLMs might push
them towards the wrong direction
talysson.com
Slide 60
Slide 60 text
Pay close attention to the confidence
provided by conversations with LLMs
talysson.com
Slide 61
Slide 61 text
What if there are no domain experts?
talysson.com
Slide 62
Slide 62 text
talysson.com
Slide 63
Slide 63 text
talysson.com
Slide 64
Slide 64 text
talysson.com
Slide 65
Slide 65 text
Use LLMs to empower your team
while the team explores the domain model
talysson.com
Slide 66
Slide 66 text
Can LLMs help during EventStorming?
talysson.com
Slide 67
Slide 67 text
Yes
But the LLM is never in the room
talysson.com
Slide 68
Slide 68 text
Roles played outside of the EventStorming room
1.
Make the industry-average knowledge useful
-
2.
3.
Generate candidate domain events before the session,
read and compare with the wall only afterwards
Role-play the absent stakeholder
-
Generate questions the absent stakeholder would make, never answers
-
Direct the questions to this person after the session
Check the graph structure
-
talysson.com
Events with no trigger or consumer, orderings that can't happen, …
Slide 69
Slide 69 text
Using LLMs to make the access to the
model easier for the whole team
talysson.com
Slide 70
Slide 70 text
LLMs don't think
You can't give them model intuition
talysson.com
Slide 71
Slide 71 text
But you can document the
model as much as possible
and feed it to a LLM
talysson.com
Slide 72
Slide 72 text
The documentation does not
contain the totality of the model
But it can be good enough!
talysson.com
Slide 73
Slide 73 text
Documenting the domain model
-
Sub-domains and bounded contexts
-
Scope and responsibilities
-
Ubiquitous language glossary
-
Context map
-
Specifications, use cases and user stories of your system
-
Tactical patterns
talysson.com
-
Aggregates, entities, value objects, domain services, …
-
Architecture Decision Record (ADR)
Slide 74
Slide 74 text
Written (or at least guided) by a human
-
Your documents should use the ubiquitous language
-
Actively guide it not to let LLMism sneak into it
-
Otherwise your docs will eventually look like this:
-
-
"This check is load-bearing in the process"
-
"And it answers the question plus expose one more wrinkle"
-
"And the trap is zero-dollar invoices"
Said no human domain expert ever
talysson.com
Slide 75
Slide 75 text
Fill the gaps where an LLM could hallucinate
-
Don't leave important concepts implicit
-
When images are needed, add a description as well
-
Explicitly state possible misunderstandings
-
E.g.: " is not the same as "
-
Zero documentation is better than misleading/outdated documentation
-
But zero documentation means the LLM gets more… creative
-
Treat hallucinations as a sign of a gap in your docs
talysson.com
Slide 76
Slide 76 text
Checking the ubiquitous language with LLMs
1.
Synonyms
-
2.
Homonyms
-
3.
"Find one word that is used for more than one concept"
Glossary drift
-
4.
"Find two or more words used for the same concept"
"Find terms in code that aren't in the glossary, and vice-versa"
Name vs behavior
-
talysson.com
"The glossary says XYZ, does the code agree?"
Slide 77
Slide 77 text
Use tools to make it even more LLM-friendly
-
Use team workspaces in LLM chat tools
-
Connect the LLM to MCP servers of your platforms
-
-
Version control system (GitHub, Gitlab, …)
-
Wiki (Confluence, Notion, …)
-
Project management (Jira, Linear, …)
Use well-known LLM-friendly formats
talysson.com
-
AGENTS.md / CLAUDE.md
-
Mermaid diagrams
-
Skills for well-defined patterns
Slide 78
Slide 78 text
talysson.com
Slide 79
Slide 79 text
LLMs as domain experts
-
LLMs can't be fully trusted, they can hallucinate
-
Arguably, humans can make mistakes too
-
The difference being that humans:
-
Can actually think
-
Have long-term memory
-
Can be held accountable
-
Domain experts provide expertise and collaboratively explore the model
-
LLMs can be used to the former, but should not make decisions on the latter
talysson.com
Slide 80
Slide 80 text
It's ok to let LLMs write the code
But the model belongs to humans
talysson.com
Slide 81
Slide 81 text
Thank you!
@talyssonoc
talysson.com