CODELEV @codelevltd

AI Action Systems for Ecommerce codelev.com Joined November 2018

Tweets

8
Followers

1
Following

7
Likes

0

CODELEV @codelevltd

4 weeks ago

Atlassian showed: 16% of SDLC time is spent on code. If a team is generating code 5x faster with AI, it will speed up the overall process only by 13%. More productive is to focus on the other 19 factors, 7 of which are about product knowledge management. #eventstorming #ai

0 0 0 4 0

View Details

CODELEV @codelevltd

2 months ago

How StormPilot embeds formal notations into Event Storming elements to reduce interpretation drift. The guide: stormpilot.codelev.com/guide

0 0 0 4 0

View Details

CODELEV @codelevltd

3 months ago

Remove the human owner of a process, and responsibility doesn't transfer to the AI that replaced them. Or does it? EU AI Act mandates human oversight only for high-risk systems. Everything below that bar runs on old-school hedges: LLCs, insurance. We underwrite the rest.

0 0 0 9 0

View Details

CODELEV @codelevltd

3 months ago

AI coding agents turn ambiguous specs into code fast. That's the problem, not the improvement. StormPilot is a Structured Event Storming editor that gets stakeholders to agree on WHAT to build before any code gets written. Demo: stormpilot.codelev.com/explore

0 0 0 10 0

View Details

CODELEV @codelevltd

4 months ago

Only 1 of the 4 signals feeding your agents is actually unambiguous. Intent: natural. Feedback: mostly natural. Environment changes: semi-formal. Formal spec notation is the only one built to remove guesswork - and it's the one signal your business side never writes. #LLM #SDLC

0 0 0 12 0

View Details

CODELEV @codelevltd

4 months ago

How do you know when LLM is guessing? Models expose logprobs - you can request the top alternative tokens and their probabilities. Low entropy = certain, high = guessing. Tested on 3 prompts 0 temperature: gpt-4o-mini: 0.00 / 0.35 / 0.55 ✓ gpt-4.1-nano: 0.39 / 0.58 / 0.35 ✗

0 0 0 16 0

View Details

CODELEV @codelevltd

5 months ago

Self-hosted LLM, temperature 0.0, same prompt. Sequential gives 1 unique answer, parallel gives 6. Batching changes floating point results enough that decoding picks a different token. Your temperature setting doesn't fix it. Run a consistency probe before you trust a model.

0 0 0 30 0

View Details

CODELEV @codelevltd

5 months ago

Accountant held her domain 2 yrs - no time for freelancers. I told Claude to build her a single-page site in light cyberpunk aesthetic. 5 mins later, done. Zero IT skills needed. AI even added a fake contact form for functionality. 😉