AI integration · Kyoto, Japan
You are not behind on AI.
You are behind on shipping it.
GKB/AI is the AI practice of GKB Studio, a one-person software studio in Kyoto run by Christopher Huber. The work is unglamorous and specific: find the part of your business where a model actually changes the arithmetic, build it properly, put it in production, and hand you something your own engineers can read. What follows are working notes — counts, entries, and one diagram — rather than claims.
this page is written to survive being forwarded to your CTO
Observations, counted
AI systems shipped to production
agents I run daily on my own machine
design to a shipped cross-platform mobile app
on the same production platform since
the 48 is the honest one — daily use is where you learn what breaks
How an engagement actually runs
- One problemworth money
- Smallest buildthat proves or kills it
- In productionmeasured, priced, owned
- Youkeep it
no discovery phase that ends in a slide deck
no discovery phase that ends in a slide deck
What this is not
Not a chatbot bolted to your website. Not a pilot that lives forever in a sandbox. Not a strategy document. The output is running software with a cost per use you can put in a spreadsheet, plus the reasoning behind every choice, written down.
What an AI system is actually made of
- 01Access to your datathe part everyone underestimatesSite specific
- 02Retrieval and context assemblydeciding what the model is allowed to seeBuilt here
- 03The prompt and tool contractwhat it can do, and what it must refuseBuilt here
- 04The modelinterchangeable, and getting cheaperBought in
- 05An eval suiteso a change that breaks it fails loudlyBuilt here
- 06Cost and rate ceilingsa bill that cannot surprise youBuilt here
- 07Retries and a fallback pathfor the day the provider is downBuilt here
- 08The human review loopwho signs, and on what evidenceBuilt with you
- 09An audit trailwhat was asked, what came back, who saw itBuilt here
One line on that list is something you buy. The other eight are the work, and they are the reason a pilot that impressed everyone in the room never made it to production — it was line 04 on its own, quietly borrowing the rest from a person who was in the demo with it.
swapping the model is a Tuesday. The other eight are the engagement.
Work, logged
- 01EntryMCP server
Your own genomics, answerable in plain language
An MCP server that lets Claude and ChatGPT query a person's own genomics and lab data directly — search it, explore it, set regimens against it. Context carries between conversations, so the assistant picks up where the last session left off instead of starting cold.
- 02EntryAI pipeline
Hours of literature reading, drafted
A multi-pass pipeline for interpreting rare genetic variants, where each pass narrows and checks the one before it. What took a scientist hours of literature reading now arrives drafted to a standard that needs only light human review.
a human still signs it — that was the point
- 03EntryAgent harness
Regulated chat you can meter and bill
A HIPAA-compliant chat harness on AWS Bedrock, built to hold coherence across long conversations and to cache prompts so cost stays flat as usage grows. File upload, PDF generation, and live token metering are in the box, so usage can be priced rather than guessed at.
The things you will ask in the first four minutes
What you own at the end
The code, the eval suite, the measured cost per use, and a written account of where the system is weak. No retainer, no licence on your own data, nothing that stops your engineers changing it the week after I leave.
If I am unavailable
One person is a real risk and pretending otherwise would be the first bad sign. It is handled with artefacts, not promises: a runbook, an eval suite that tells your team when something has broken, an architecture note, and a written list of the failure modes I already know about.
How the counts above are meant
They are counts, not outcomes. There is no accuracy percentage, no growth curve and no saving-per-year on this page, because I have not measured those things here and drawing them would be the easiest lie on the internet. Ask on a call and you will get the figures that do exist, with their caveats.
Where the six came from
One client, under NDA — a consumer genomics and lab-testing platform I have worked inside since 2021, as an employee and then as a retained client. The AI systems on it date from 2024 onwards, and all six are still in production. Six shallow engagements would read better and mean less: four years in one codebase is where you find out what an AI system does in its second year.
Next action
Bring the one process you suspect a model could take over.
One conversation is usually enough to say whether it is worth building, worth waiting on, or not an AI problem at all. If the honest answer is the third one, you will hear that.
And if you already know exactly what you want built, skip all of this and email me.
replies come from Christopher, not a form