Blog

Day N

A development diary — building SaaS products solo with AI. Real decisions, real mistakes, real lessons.

North StarStart here2026-09-30 · 5 min readAI Governance

I Asked for "Another Me." Here's What It Took to Trust It.

What I told Anthropic I want from AI: a partner that remembers my work with evidence. Memory turned out to be the easy part. Trust was the work, and every piece of it came from a failure.

Counts come from the published essay metadata, one essay per tag. An essay can appear in several groups. As of the latest listed editorial revision: 2026-10-01.

Group by

Build in Public

7
2026-09-06 · 7 minAI Agents

Before Explaining the Spike, I Had to Check the Question

A Grand Prize OpenSearch hackathon build, shaped by mistakes in a skating app and a learning app: check the incident's premise, compare each group with its own baseline, and leave the missing evidence visible.

2026-09-06 · 6 minAI Agents

The Catalog Was Updated. The Data Was Still from 2016.

A fresh catalog timestamp was answering the wrong question. Building a DataHub agent that compares declared cadence with data time, names what it cannot measure, and checks that its findings actually reached the catalog.

2026-09-06 · 6 minIvyBloom

My Science Questions Passed Every Check. They Mostly Tested Subtraction.

A question can depend on its data table and still need none of the science it is meant to teach. What an IvyBloom review exposed about the gap between a passing criterion and the actual learning task.

2026-08-24 · 5 minAI Agents

A Spreadsheet Does Not Look Different When It Is Lying

A Google All Things Agentic Hackathon build story: a reconciliation agent for homeschool instruction-time evidence that attests only what the evidence can back — and treats refusal as an autonomous action.

2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.

2026-06-26 · 4 minJumpOnion

When My Figure-Skating AI Called a World Champion's Triple Axel "Under-Rotated"

During development I pointed JumpOnion's rotation detector at a world champion's textbook triple axel. It flagged under-rotation — and that single false positive decided how I built the rest of the product.

2026-06-26 · 3 minIvyBloom

One Person, Two Kids, and 140,000 Lines of Code

I'm one person. With AI I built and still maintain a learning app for my two kids — 85 API routes and around 140,000 lines of code. Here's what "AI-built" actually means when your entire user base is two people you see at breakfast.

AI Engineering

6
2026-09-06 · 6 minIvyBloom

My Science Questions Passed Every Check. They Mostly Tested Subtraction.

A question can depend on its data table and still need none of the science it is meant to teach. What an IvyBloom review exposed about the gap between a passing criterion and the actual learning task.

2026-04-30 · 8 minJumpOnion

Day 8: The LLM Succeeded. The Platform Killed the Connection.

First production smoke of a new training-plan endpoint. The browser hung, returned 504, and a manual refresh showed the plan was already there. The LLM finished; the platform gateway killed the client first.

2026-04-22 · 7 minJumpOnion

Day 6: A Button Name Killed My Algorithm

A month of algorithm tuning hit clean calibration on real videos. Then a parent uploaded one, and the system reported an impossible air-time reading. The algorithm was right. The button label was the bug.

2026-04-14 · 8 minJumpOnion

Day 3: When AI Lied About Dead Code (30 Hours of Silent Failures)

My AI partner deleted production code with absolute confidence and labeled it 'dead code, anonymous uploads.' 987 tests passed. The deploy went green. Thirty hours later, a paying customer noticed.

2026-04-13 · 7 minJumpOnion

Day 2: 6.47 Seconds of Air Time (The Jump Lasted 0.7)

The first attempt at detecting a figure skating jump measured 6.47 seconds of air time — on a jump that lasted 0.7 seconds. Three failed approaches, then the system rescued itself.

2026-04-08 · 6 minJumpOnion

Day 1: Building a Figure Skating AI — Why I Started

My kids train figure skating several days a week. Coaches cost $120/hour and their time is limited. I built an AI that diagnoses jumps in 60 seconds — here's why I started and what the first day looked like.

Verification

6
2026-10-01 · 9 minOpenSearch

My Agents Told Me to Enter. Here's What Happened at OpenSearchCon.

Before I went on stage at OpenSearchCon, I told the organizers that my own agents had told me this hackathon was right for me. A short film and the story behind it: building unclosed, three days among people who run search at enormous scale, and a trust layer I brought home.

2026-09-06 · 7 minAI Agents

Before Explaining the Spike, I Had to Check the Question

A Grand Prize OpenSearch hackathon build, shaped by mistakes in a skating app and a learning app: check the incident's premise, compare each group with its own baseline, and leave the missing evidence visible.

2026-09-06 · 6 minAI Agents

The Catalog Was Updated. The Data Was Still from 2016.

A fresh catalog timestamp was answering the wrong question. Building a DataHub agent that compares declared cadence with data time, names what it cannot measure, and checks that its findings actually reached the catalog.

2026-09-06 · 6 minIvyBloom

My Science Questions Passed Every Check. They Mostly Tested Subtraction.

A question can depend on its data table and still need none of the science it is meant to teach. What an IvyBloom review exposed about the gap between a passing criterion and the actual learning task.

2026-08-24 · 5 minAI Agents

A Spreadsheet Does Not Look Different When It Is Lying

A Google All Things Agentic Hackathon build story: a reconciliation agent for homeschool instruction-time evidence that attests only what the evidence can back — and treats refusal as an autonomous action.

2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.

Hackathon

5
2026-10-01 · 9 minOpenSearch

My Agents Told Me to Enter. Here's What Happened at OpenSearchCon.

Before I went on stage at OpenSearchCon, I told the organizers that my own agents had told me this hackathon was right for me. A short film and the story behind it: building unclosed, three days among people who run search at enormous scale, and a trust layer I brought home.

2026-09-06 · 7 minAI Agents

Before Explaining the Spike, I Had to Check the Question

A Grand Prize OpenSearch hackathon build, shaped by mistakes in a skating app and a learning app: check the incident's premise, compare each group with its own baseline, and leave the missing evidence visible.

2026-09-06 · 6 minAI Agents

The Catalog Was Updated. The Data Was Still from 2016.

A fresh catalog timestamp was answering the wrong question. Building a DataHub agent that compares declared cadence with data time, names what it cannot measure, and checks that its findings actually reached the catalog.

2026-08-24 · 5 minAI Agents

A Spreadsheet Does Not Look Different When It Is Lying

A Google All Things Agentic Hackathon build story: a reconciliation agent for homeschool instruction-time evidence that attests only what the evidence can back — and treats refusal as an autonomous action.

2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.

AI Agents

4
2026-10-01 · 9 minOpenSearch

My Agents Told Me to Enter. Here's What Happened at OpenSearchCon.

Before I went on stage at OpenSearchCon, I told the organizers that my own agents had told me this hackathon was right for me. A short film and the story behind it: building unclosed, three days among people who run search at enormous scale, and a trust layer I brought home.

2026-09-06 · 7 minAI Agents

Before Explaining the Spike, I Had to Check the Question

A Grand Prize OpenSearch hackathon build, shaped by mistakes in a skating app and a learning app: check the incident's premise, compare each group with its own baseline, and leave the missing evidence visible.

2026-09-06 · 6 minAI Agents

The Catalog Was Updated. The Data Was Still from 2016.

A fresh catalog timestamp was answering the wrong question. Building a DataHub agent that compares declared cadence with data time, names what it cannot measure, and checks that its findings actually reached the catalog.

2026-08-24 · 5 minAI Agents

A Spreadsheet Does Not Look Different When It Is Lying

A Google All Things Agentic Hackathon build story: a reconciliation agent for homeschool instruction-time evidence that attests only what the evidence can back — and treats refusal as an autonomous action.

AI Governance

3
2026-09-30 · 5 minAI Governance

I Asked for "Another Me." Here's What It Took to Trust It.

What I told Anthropic I want from AI: a partner that remembers my work with evidence. Memory turned out to be the easy part. Trust was the work, and every piece of it came from a failure.

2026-05-21 · 9 minIvyBloom

Day 9: Your AI Debug Log Isn't Private

A grep over the knowledge vault hit a three-month-old AI session archive containing a full .env dump. The vault syncs to the cloud. The cloud indexes into AI search. A private debug log had become a three-layer attack surface.

2026-04-15 · 8 minJumpOnion

Day 4: Memory Is a Wish. Hooks Are the Law.

The anti-pattern was already in memory. The AI loaded it at session start. It walked right into the same bug anyway. Forty minutes from paying-customer report to root cause to deploy — and a deeper finding that rewired how I store rules.

Memory Architecture

2
2026-09-30 · 5 minAI Governance

I Asked for "Another Me." Here's What It Took to Trust It.

What I told Anthropic I want from AI: a partner that remembers my work with evidence. Memory turned out to be the easy part. Trust was the work, and every piece of it came from a failure.

2026-04-15 · 8 minJumpOnion

Day 4: Memory Is a Wish. Hooks Are the Law.

The anti-pattern was already in memory. The AI loaded it at session start. It walked right into the same bug anyway. Forty minutes from paying-customer report to root cause to deploy — and a deeper finding that rewired how I store rules.

OpenSearch

2
2026-10-01 · 9 minOpenSearch

My Agents Told Me to Enter. Here's What Happened at OpenSearchCon.

Before I went on stage at OpenSearchCon, I told the organizers that my own agents had told me this hackathon was right for me. A short film and the story behind it: building unclosed, three days among people who run search at enormous scale, and a trust layer I brought home.

2026-09-06 · 7 minAI Agents

Before Explaining the Spike, I Had to Check the Question

A Grand Prize OpenSearch hackathon build, shaped by mistakes in a skating app and a learning app: check the incident's premise, compare each group with its own baseline, and leave the missing evidence visible.

AI Built

1
2026-06-26 · 3 minIvyBloom

One Person, Two Kids, and 140,000 Lines of Code

I'm one person. With AI I built and still maintain a learning app for my two kids — 85 API routes and around 140,000 lines of code. Here's what "AI-built" actually means when your entire user base is two people you see at breakfast.

AI Hiring

1
2026-06-05 · 9 minEnterprise AI

"AI Engineer" Is a Fake Job Title

A company decides it needs AI, hires an expensive 'AI Engineer,' and stalls at month three. Then it hires someone pricier — and stalls again. The logic of the hire was broken at the root — and the role that actually works has no market label yet, because it can only come from inside.

AI Quoting

1
2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.

AI Tooling

1
2026-04-18 · 9 minSOP Framework

Day 5: 89K Stars. I Adopted 3 Ideas. Rejected 3.

An open-source AI agent framework went viral promising self-evolution through genetic mutation. The temptation was full adoption. The structured 5-dimension comparison was cheaper — and gave me 3 patterns to steal and 3 to reject on sight.

Calibration

1
2026-06-26 · 4 minJumpOnion

When My Figure-Skating AI Called a World Champion's Triple Axel "Under-Rotated"

During development I pointed JumpOnion's rotation detector at a world champion's textbook triple axel. It flagged under-rotation — and that single false positive decided how I built the rest of the product.

Competitive Analysis

1
2026-04-18 · 9 minSOP Framework

Day 5: 89K Stars. I Adopted 3 Ideas. Rejected 3.

An open-source AI agent framework went viral promising self-evolution through genetic mutation. The temptation was full adoption. The structured 5-dimension comparison was cheaper — and gave me 3 patterns to steal and 3 to reject on sight.

Data Safety

1
2026-04-25 · 9 minJumpOnion

Day 7: When AI Wiped Generated Outputs Across 20 Rows (in One SQL Statement)

After hunting down a real production blocker, the AI moved to clean up a stale cache. The SQL nullified three independent caches in one go. 20 rows updated. 7 paying customers' personalized diagnoses and training plans — gone.

DataHub

1
2026-09-06 · 6 minAI Agents

The Catalog Was Updated. The Data Was Still from 2016.

A fresh catalog timestamp was answering the wrong question. Building a DataHub agent that compares declared cadence with data time, names what it cannot measure, and checks that its findings actually reached the catalog.

Debugging

1
2026-04-13 · 7 minJumpOnion

Day 2: 6.47 Seconds of Air Time (The Jump Lasted 0.7)

The first attempt at detecting a figure skating jump measured 6.47 seconds of air time — on a jump that lasted 0.7 seconds. Three failed approaches, then the system rescued itself.

Edtech

1
2026-06-26 · 3 minIvyBloom

One Person, Two Kids, and 140,000 Lines of Code

I'm one person. With AI I built and still maintain a learning app for my two kids — 85 API routes and around 140,000 lines of code. Here's what "AI-built" actually means when your entire user base is two people you see at breakfast.

Enterprise AI

1
2026-06-05 · 9 minEnterprise AI

"AI Engineer" Is a Fake Job Title

A company decides it needs AI, hires an expensive 'AI Engineer,' and stalls at month three. Then it hires someone pricier — and stalls again. The logic of the hire was broken at the root — and the role that actually works has no market label yet, because it can only come from inside.

Figure Skating AI

1
2026-06-26 · 4 minJumpOnion

When My Figure-Skating AI Called a World Champion's Triple Axel "Under-Rotated"

During development I pointed JumpOnion's rotation detector at a world champion's textbook triple axel. It flagged under-rotation — and that single false positive decided how I built the rest of the product.

Future of Work

1
2026-06-05 · 9 minEnterprise AI

"AI Engineer" Is a Fake Job Title

A company decides it needs AI, hires an expensive 'AI Engineer,' and stalls at month three. Then it hires someone pricier — and stalls again. The logic of the hire was broken at the root — and the role that actually works has no market label yet, because it can only come from inside.

Google Cloud

1
2026-08-24 · 5 minAI Agents

A Spreadsheet Does Not Look Different When It Is Lying

A Google All Things Agentic Hackathon build story: a reconciliation agent for homeschool instruction-time evidence that attests only what the evidence can back — and treats refusal as an autonomous action.

Infrastructure

1
2026-04-30 · 8 minJumpOnion

Day 8: The LLM Succeeded. The Platform Killed the Connection.

First production smoke of a new training-plan endpoint. The browser hung, returned 504, and a manual refresh showed the plan was already there. The LLM finished; the platform gateway killed the client first.

Post Mortem

1
2026-04-14 · 8 minJumpOnion

Day 3: When AI Lied About Dead Code (30 Hours of Silent Failures)

My AI partner deleted production code with absolute confidence and labeled it 'dead code, anonymous uploads.' 987 tests passed. The deploy went green. Thirty hours later, a paying customer noticed.

Production Incident

1
2026-04-25 · 9 minJumpOnion

Day 7: When AI Wiped Generated Outputs Across 20 Rows (in One SQL Statement)

After hunting down a real production blocker, the AI moved to clean up a stale cache. The SQL nullified three independent caches in one go. 20 rows updated. 7 paying customers' personalized diagnoses and training plans — gone.

QwenCloud

1
2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.

Security

1
2026-05-21 · 9 minIvyBloom

Day 9: Your AI Debug Log Isn't Private

A grep over the knowledge vault hit a three-month-old AI session archive containing a full .env dump. The vault syncs to the cloud. The cloud indexes into AI search. A private debug log had become a three-layer attack surface.

Solo Founder

1
2026-06-26 · 3 minIvyBloom

One Person, Two Kids, and 140,000 Lines of Code

I'm one person. With AI I built and still maintain a learning app for my two kids — 85 API routes and around 140,000 lines of code. Here's what "AI-built" actually means when your entire user base is two people you see at breakfast.

SOP Framework

1
2026-04-18 · 9 minSOP Framework

Day 5: 89K Stars. I Adopted 3 Ideas. Rejected 3.

An open-source AI agent framework went viral promising self-evolution through genetic mutation. The temptation was full adoption. The structured 5-dimension comparison was cheaper — and gave me 3 patterns to steal and 3 to reject on sight.

Trust

1
2026-09-30 · 5 minAI Governance

I Asked for "Another Me." Here's What It Took to Trust It.

What I told Anthropic I want from AI: a partner that remembers my work with evidence. Memory turned out to be the easy part. Trust was the work, and every piece of it came from a failure.

UX Design

1
2026-04-22 · 7 minJumpOnion

Day 6: A Button Name Killed My Algorithm

A month of algorithm tuning hit clean calibration on real videos. Then a parent uploaded one, and the system reported an impossible air-time reading. The algorithm was right. The button label was the bug.

Vertical AI

1
2026-07-07 · 6 minAI Quoting

The Trades Are Raising Billions for AI. The Quote Is Still the Hard Part.

A QwenCloud Global AI Hackathon Track 4 build story: an AI quoting agent where Qwen handles negotiation rehearsal, database truth stays deterministic, and humans approve where money moves.