Two AI builds by Fabien Starseed with traceable numbers: an ebook pipeline whose review gate removed 143 unsourced claims, and a go or no-go weather cockpit for offshore wind inspection crews.
Every number on this page comes from a review log, a code file or a dated work log, cited next to it. Client projects stay confidential, so no client names, photos, farms or operators appear here.
Own product · EbookForge · Gemini 2.5 Flash, Next.js, AI review panel · Nov 2026
143 unsourced claims removed before any ebook reached a reader
The problem
AI drafts read well and still invent things. The first drafts of my field-guide ebooks contained invented command-line flags, wrong definitions of hooks and MCP, and unsafe operational numbers. Publishing them as they were would have put false instructions in front of paying readers.
What I built
A pipeline that drafts each chapter with Gemini 2.5 Flash, then sends the whole text to a review panel before any cover or PDF exists: five reviewers (aviation and maritime counsel, copyright and plagiarism with real web searches, a fact auditor, trademark and consumer protection, a red team) and a lead counsel who rules PASS, FIX or BLOCK. The order is enforced in code: the cover and PDF steps refuse to run without a verify stamp.
The numbers
Drafts reviewed in round 1 (4 Nov 2026)
3, all sent back as FIX
Source Review logs, 4 Nov 2026
Claims deleted in round 1
143 (60 + 23 + 60)
Source Review logs, 4 Nov 2026
Claims confirmed against a source in round 1
18 (11 + 7 + 0)
Source Review logs, 4 Nov 2026
Plagiarism web searches in round 1
147 (73 + 32 + 42)
Source Review logs, 4 Nov 2026
Exact-phrase plagiarism searches in round 2 (10 Nov 2026, two books)
169, with 0 copied passages found
Source Review logs, 10 Nov 2026
Time to draft one chapter after the model switch
About 18 s instead of about 90 s (5x faster)
Source Build notes, model benchmark
Where it stands
No ebook has passed the gate yet, so none is on sale. That is the gate doing its job: nothing reaches a reader until every claim in it can be sourced.
What it means for your business
The same pattern works for proposals, reports and customer answers: one AI step produces the draft, a separate checker deletes what it cannot prove, and nothing goes out without the checker's stamp.
Own product · Wind Hub Weather · Google DeepMind WeatherNext, Open-Meteo, Next.js · Sep to Nov 2026
A go or no-go weather call for offshore inspection crews
The problem
On an offshore wind farm, wind, waves, swell and visibility decide whether a crew transfer vessel and a drone team can work that day. Captains and pilots need one clear answer for the day.
What I built
A weather view that merges Google DeepMind WeatherNext with Open-Meteo hour by hour and scores every hour against the limits of an offshore blade-inspection day. It is honest about what it does not know: waves are forecast 7 days ahead only, and later days are capped at caution instead of guessing.
The numbers
Wind gust limits
Go up to 8 m/s, caution up to 12 m/s
Source Wind Hub source code
Significant wave height limits
Go up to 1.2 m, caution up to 1.8 m
Source Wind Hub source code
Swell period and visibility
Caution from 8 s, no-go from 11 s; caution under 2,000 m visibility
Source Wind Hub source code
Wave forecast horizon
7 days, later days capped at caution
Source Wind Hub source code
Captain feedback turned into a redesign (22 Oct 2026)
6 quick spot buttons replaced tapping on the map
Source Work log, 22 Oct 2026
Free WeatherNext forecasts per user per day
3
Source Wind Hub source code
Where it stands
Crew transfer vessel captains are testing the weather view, and their feedback on LINE drives each redesign. Their first note was too much information and a map that was hard to tap, so the main view now shows wave, wind and gust tiles first. No usage counts are published yet.
What it means for your business
The same approach fits any go or no-go decision built on several data sources: merge them, score them against limits your team already uses, and show one answer with its reasons, including what the data cannot tell you.