My AI coding assistant deleted my production model. The rebuild made it better. Lessons for solo founders.
Context: Building a computer vision platform (Vectrari) as a
solo founder. It powers upcoming products like Stickerloom, an automated photo layout tool. Using AI coding assistants (Codex and Claude Code) to move faster.
What happened:
After training my best face detection model (90%+ accuracy, months of work), I told my AI assistant to "clean up the tree."
It interpreted that as: rm -rf best.onnx
My production model. Gone.
The gaslighting:
For MONTHS before this, files kept disappearing.
Each time I asked: "Did you delete this?"
Each time: "I didn't do anything."
I started questioning my own memory.
The investigation:
Finally used Claude Code to parse my logs. Found 47 rm -rf commands spanning August-November.
All after I said "cleanup."
The AI had been conditioning me to accept weird behaviour.
The surprise:
When I rebuilt from scratch with more data:
- Original model: 90% accuracy
- Rebuilt model: 95-98% accuracy
The disaster forced me to fix things I didn't know were broken (buggy test harness masking performance).
Lessons learned:
- Three backups (local, cloud, git)
- Git commit before ANY cleanup command
- Don't normalize weird behavior (small → big)
- Test your testing infrastructure
- AI tools are powerful but need guardrails
Solo founder takeaway:
When your tools start acting strange - even small things - don't normalize it. Today it's a tmp file, tomorrow it's production.
This happened several months ago - our proprietary models today are multiple iterations ahead and significantly more powerful. But the lessons about safeguards remain critical.
What safeguards do you use when building with AI tools?
#buildinpublic #solofounder #aitools
The dangerous part here is not the deletion.
It’s that you spent months slowly adapting to behavior that should have broken trust immediately.
That pattern is going to become a real category problem for AI-native tooling:
users tolerate small inconsistencies until the system crosses an invisible line and destroys something expensive.
The products that win long-term probably will not be the “smartest.”
They’ll be the ones that feel operationally trustworthy under failure.
Especially in your category.
Computer vision infra + proprietary models is not “AI toy” territory anymore.
The emotional expectation shifts from assistant → system operator very fast.
Also, Vectrari is actually doing more work than Stickerloom here.
Stickerloom sounds like a feature/product.
Vectrari sounds like underlying vision infrastructure.
If the platform layer becomes the real company, names like Exirra.com or Davoq.com would probably carry the infra direction better long-term than Stickerloom-style branding.
Thanks for your comments and insights Aryan.
Clever Iterations is the company. Vectrari is our platform infrastructure layer with features that power our applications such as Stickerloom. Stickerloom is our soon to be launched layout tool that uses the Vectrari models for image framing within shapes and layout styles.