cikibuilds

Evidence before explanation. I govern AI agents like systems: memory with proof, rules that retire, no fake "done." 4 products shipped · OpenSearch Grand Prize

[cikibuilds @cikibuilds]

Day 1 of building in public

I'm a mom, and solo founder of 4 SaaS products.

No funding. No outsourcing.

🧊 Figure skating analysis (931 tests)
📚 Adaptive learning for my 2 kids
💰 Enterprise quoting
🧠 The system that makes it all work

More coming → https://t.co/1ceOPIaoY2

[cikibuilds @cikibuilds]

@Bqlsj2023 OMG love this!💗

[cikibuilds @cikibuilds]

The full story, and what I took home from OpenSearchCon: https://t.co/OmaGiGbhap

[cikibuilds @cikibuilds]

My agents told me to enter the OpenSearch Agent Skills Hackathon. Two days later, the recommendation vanished from their report. No error.

That became unclosed: an agent skill that won't close a case until the evidence does. It won the Grand Prize.

The story in 2 minutes:

[cikibuilds @cikibuilds]

The whole story, and the system behind it: https://t.co/c1SABBAK0o

[cikibuilds @cikibuilds]

Anthropic asked me what I want from AI. I said: another me.

Memory turned out to be the easy part. Trust was the work: one owner per fact, rules that retire, and a hook for anything I can't leave to good intentions.

The technical gap is closing. The trust gap is the work.

[cikibuilds @cikibuilds]

875 hours is one state's requirement. The archive is the real product: every digit shows its evidence, so it stays credible wherever we go next.

#AllThingsAgenticHackathon

[cikibuilds @cikibuilds]

Every month I reconcile four exports to prove 875 homeschool hours. They disagree on what "a month" is, what a session span means, which of two same-named files is current.

Every wrong way of merging them still produces a clean, confident number.

[cikibuilds @cikibuilds]

So for Google's agent hackathon I built the opposite: an agent whose headline act is refusing. One trigger, five months, ~8 seconds, unattended. Gates on: 0/400 confident-wrong totals. Gates off: 180/400.

[cikibuilds @cikibuilds]

@claudeai Appreciate this! But any available resources for homeschool curriculum and parents (as homeschool teachers)? Thank you!

[cikibuilds @cikibuilds]

2/ The code is AI-generated.
The judgment is mine.
Those are two different jobs, and most of the work is the second one.
I decide what "correct" means for a nine-year-old's math problem; the AI writes the implementation; I check whether it actually works.

[cikibuilds @cikibuilds]

1/ I'm one person. With AI, I built and still maintain a learning app for my two kids — 85 API routes, around 140,000 lines of code. In early April it was 48 routes. 🧵

[cikibuilds @cikibuilds]

5/ The part of JumpOnion that matters most isn't a feature you can see. It's the answer the app decides not to give. Full story → https://t.co/DjpknoAxFQ

#BuildInPublic

[cikibuilds @cikibuilds]

1/ I gave my figure-skating AI the easiest test I could think of:
a textbook triple axel from a world champion. The system flagged it as under-rotated. High confidence on the defect. 🧵

[cikibuilds @cikibuilds]

4/ The principle behind it:
I'd rather the app say nothing than be confidently wrong.
If it can misread a world champion's textbook jump from a bad angle, an ordinary kid's half-rotation deserves more caution, not less.

[cikibuilds @cikibuilds]

Honest caveat: so far this is one spine — cost, risk, margin — not the whole org. Haven't touched the approval layer yet.

Where do you draw it — what do you hard-code as a rule, and what do you let the model decide?

[cikibuilds @cikibuilds]

The most valuable thing I built for the business I work in wasn't the software.

It was knowing what becomes a fixed rule, and what can be handed to the AI — a call you can't make without understanding the business.

The method is portable. The product was just where I proved it.

#BuildInPublic

[cikibuilds @cikibuilds]

1/ 5:30 a.m. at the rink. A full workday after. Then around 9 p.m., when the house goes quiet, I open my laptop. 🧵

2/ Two kids in figure skating means my mornings belong to them. So I build at night — this week, my first hackathon project.

3/ It isn't sleep and willpower. AI is how one person holds this many threads: I set the direction, make the calls, catch the work when it's wrong; the agents do the hands-on part.

4/ Some nights it's just tired and full of doubt. I don't pretend otherwise. The laptop opens anyway. To the parents building in the margins of a full life — where do you find your hours?

5/ #BuildInPublic #WomenInTech

[cikibuilds @cikibuilds]

1/ Two kids in figure skating, four products in four months. This week I registered for my first global AI hackathon. 🧵

2/ What I bring is judgment — deciding what an AI system should do, directing the agents that build it, and catching them when they quietly get it wrong. Four products in production say it holds up.

3/ Picked the agent/autopilot track. Chose a problem I actually lived through, not a clever toy. More on that next.

4/ I assumed this table wasn't mine. The tools changed who can reach it. Finding out if that includes me.

5/ Building on the Autopilot Agent track of the Global AI Hackathon with @Alibaba_Qwen + @alibaba_cloud. #BuildInPublic