Your posts have another version.

Our design system looked organized.

To an agent, it was chaos.

For a couple of months we were using one giant folder and letting AI design new features from the design system

The problem was simple. The system was one big folder with everything in it. Components, patterns, styles and variants were all exposed at once.

That gave the agent too many choices.

The output was often close but not clearly aligned with the rest of the product. It could generate UI but the decisions were not deterministic enough.

So we changed the structure.

We built a knowledge graph on top of the design system.

Instead of giving the agent raw access to everything, we mapped components, variants, colors, patterns, examples, business context and the rules for how each part connects.

Then we plugged that into the agent workflow through MCP.

That is when the output became much clearer and much better aligned with existing features.

Same AI.
Better system.

I think this is the real shift with a context layer.

Agent-savvy design systems will matter more than just giving agents access to a big component folder.

Is anyone else rebuilding their design system this way?

#design

In legal AI RAG finds the clause. Context graphs find who it binds.

Ive been working on memory layers for legal AI for quite some time now. One thing I keep noticing is that most companies use the same RAG pipeline for every kind of problem.

Chunk. Embed. Retrieve. Hope.

It works for finding a clause. It breaks when the answer depends on how facts connect. The retriever pulls passages that look relevant and misses the ones that link them. The model never sees the connection so it never extracts the right fact.

So I tested it on a real shape of problem.
2,000 documents from a merger of two companies.

Company A alone has many entities across subsidiaries and affiliates and third parties. Each one is tied to contracts. Each contract is tied to parties and locations.

Now ask a simple diligence question.
Which third parties does Company A have agreements with? Through which entity?

RAG answers from whichever chunks rank in the top k. If the contract that links an entity to a third party isnt in those chunks it doesnt exist for the model.

Heres the pipeline I built instead
parse_file → classify_contract → extract_affiliate or extract_co_branding → build_graph

Classification runs first because every contract type carries different obligations. Each type gets its own extractor that knows what to pull.

Contract → HAS_PARTY → Entity
Entity → HAS_LOCATION → Address

Start at Company A. Walk every entity. Walk every contract. Nothing depends on ranking.

If youre building in legal AI how are you handling the context layer? Still pure RAG or are you modeling relationships too?

#legalai #contextgrpah

I helped a friend's startup plan an AI-native tech team structure and the first thing I dropped from the hiring bar was language and framework questions.

With agentic AI advancing this fast, it amazes me how many companies still interview the old way. Write this design pattern from memory. Get every bit of syntax right on a whiteboard. Explain how that framework handles X.

That's the exact work the same team will hand to AI on day one.

So this is how I see the bar at each level for AI-native roles.

STAFF, LEAD AND PRINCIPAL
- Sets technical direction and owns key architecture decisions
- Drives AI integration while keeping oversight of every change and decision
- Manages multiple teams and helps them navigate tough tech problems

MID TO SENIOR
- Asks how, why and what we are doing
- Works with product managers and designers to think through ideas and ship features
- Connects the dots and backs ideas with metrics
- Builds and contributes to the structure created by staff
- Seniors unblock the team and keep work moving

JUNIOR AND FRESHERS
- Fundamentals over any specific language
- A couple of projects built on anything
- Curious and knows when to ask for help

What's one thing you've changed in how you hire engineers since AI joined the team?

#hiring #tech

Got a new customer for one of my HRtech B2B apps this month. Never talked to them once.

With AI in SaaS the freemium debate is back again. Is it worth it or not as with AI inference cost is high?

This customer was on the free tier for 2 months. Used it with 25 users. Hit the limit and upgraded on their own.

What matters with freemium is simple. Can someone get the value without you there?

For this customer the answer was yes. They tried it in their workflow and by the time they hit the cap they already knew it worked.

That is what freemium did for me. It cut the explanation and let the product speak first.

For marketplace SaaS that can work really well.

The harder part is the free users who stay for months. They get enough value to keep using it. Not enough value to pay.

That is the part I still have not cracked 🙃