Ryan Tsuji

CTO @airCloset (1.4M users). Built an Agentic Graph RAG over our codebase + DBs. Writing about AI infra that actually works in production. Tokyo 🇯🇵

[Ryan Tsuji @ryantsuji]

Following up on the AI harness post from last week: here's the heart of it.

I redefined our entire dev platform with one premise — built for AI, operated by AI, used by AI. This post is about the knowledge graph that makes that possible.

https://t.co/f2PzyYsnC5

[Ryan Tsuji @ryantsuji]

"AI changes everything," people keep saying. After building and running an AI platform for our org, my biggest takeaway is the opposite.

The principles don't change. Only the machinery does.

I wrote up the thinking behind everything I've published so far, from code graphs to AI observability to automated review.
https://t.co/WZunZxiC7Q

[Ryan Tsuji @ryantsuji]

Part 2 of the observability series is out, the practice side.

The focus is PII, the part whose weight has gone up most in the AI era: hashing it so plain text never enters the stack, letting AI search logs without ever seeing personal data, and driving self-healing on top of that foundation.

If Part 1 was about shaping telemetry for AI, this one is about doing it safely.

https://t.co/G0MITxVImR

[Ryan Tsuji @ryantsuji]

Observability Design for the AI Era (Part 1, the design side).

"AI era" in the title, but the core question hasn't changed: why you want to measure something, and how you shape it once you do. It's about the design judgments behind that, not a new stack.

Four surfaces (app / infra / CI / LLM), each shaped to the question it answers.

https://t.co/QNQDG4B8Aw

[Ryan Tsuji @ryantsuji]

Last week: building a knowledge graph across 46 production repos via static analysis. Without semantic search though, only half-useful.

Part 2 is up — how I gave it context. ~50k MCP tool calls / 73 users in 2.5 months, including stylists & support, not just engineers.

https://t.co/3owiIKqkKQ

[Ryan Tsuji @ryantsuji]

@DorMoshe @SylwiaLask8977 @ThePracticalDev @dzakh_dev @ITNEXT_io @ctrotechDev Thank you for reading my article deeply and sharing!

[Dor Moshe @DorMoshe]

📰👨‍💻 I just published #Angular #TypeScript Trends #425 https://t.co/mXfjValj9G @SylwiaLask8977 @thepracticaldev @ryantsuji @dzakh_dev @itnext_io @ctrotechDev @dormoshe #dormosheio #javascript #frontend #webdev

[Ryan Tsuji @ryantsuji]

My post made @ThePracticalDev's Top 7 Featured DEV Posts this week 🎉

It's the series finale on the philosophy behind cortex (our internal AI platform) and how I design harnesses around AI rather than trust it.

DEV's list: https://t.co/D0hbq3680B
Article: https://t.co/wNGw4zNqLA

[Ryan Tsuji @ryantsuji]

46 repos of legacy production code unified into one knowledge graph via static analysis.

The static-analysis layer alone wasn't enough — and honestly, building it was hell 😅

If you're stuck somewhere similar, hopefully this helps.
https://t.co/bGDRJCldD0

[Ryan Tsuji @ryantsuji]

Series Final is out. After six posts on the AI harness I built, the throughline came down to one line:

"AI isn't something to trust — it's something to design."

https://t.co/87ztrIpG00

[Ryan Tsuji @ryantsuji]

Part 5 ("The Author Doesn't Have to Be an Engineer") has been generating sharp comments on https://t.co/tB4Uso717B:

"Standing up a stack is hard, building on top of one isn't."
"The harness translates domain taste into checks the system can enforce."
"Who lays the rails in ten years?"

Worth a read for the thread alone.
https://t.co/NFkH2ERIsc

[Ryan Tsuji @ryantsuji]

Part 5 of the AI harness series is out.
Non-engineers (PMOs, business leads) ship PRs straight to the production repo. The harness owns quality. Real PR examples of what actually merges without an engineer touching it.
https://t.co/VrJzRTTxk9

[Ryan Tsuji @ryantsuji]

@devlucas56 Not sure I follow — what do you mean?

[Ryan Tsuji @ryantsuji]

Following up on the AI harness post from last week: here's the heart of it.

I redefined our entire dev platform with one premise — built for AI, operated by AI, used by AI. This post is about the knowledge graph that makes that possible.

https://t.co/f2PzyYsnC5

[Ryan Tsuji @ryantsuji]

@devlucas56 AI reviews PRs, fixes request changes, self-healing, recurrence prevention and enable adding features without engineers. These are series articles. https://t.co/Ahg1VEurjv

[Ryan Tsuji @ryantsuji]

It's interesting 20% of the access to my https://t.co/sIHql5Hfdu articles is from chatgpt.

[Ryan Tsuji @ryantsuji]

Result: the pattern that caused the incident becomes structurally unwritable from then on.

Self-healing fixes symptoms. Self-healing + self-strengthening is what makes the quality gate compound over time.

[Ryan Tsuji @ryantsuji]

For Part 4 of the AI Harness series I just published: "Fixed Before Anyone Notices, Stronger After Every Fix — Self-Healing + Recurrence Prevention". The pitch is less about the auto-fix part, more about how every fix has to add a new Guide.

https://t.co/QSWsVKeo0l

[Ryan Tsuji @ryantsuji]

But "just fix it" doesn't compound.

cortex's rule: every fix PR must add one new Guide.
- lint rule
- CI guard
- type constraint
- guideline entry

Without it, the AI reviewer issues REQUEST_CHANGES — can't merge.

[Ryan Tsuji @ryantsuji]

The role shifts from human-in-the-loop to human-on-the-loop. Engineers don't write or review code at this point — they make judgment calls.

The review guidelines the AI follows (JP/EN) are open source:
https://t.co/eV2hzW7hCL

[Ryan Tsuji @ryantsuji]

Part 3 of the AI Harness series is up.

People say AI-assisted dev hits a review bottleneck and quality drops. In practice we see the opposite: review velocity and quality both go up, but only if the harness extends to the reviewer side too.

https://t.co/NxHFFJgVii

[Ryan Tsuji @ryantsuji]

Following up on the AI harness post from last week: here's the heart of it.

I redefined our entire dev platform with one premise — built for AI, operated by AI, used by AI. This post is about the knowledge graph that makes that possible.

https://t.co/f2PzyYsnC5