George Bloor

Furnace engineer running a real engineering firm on AI agents, and building B2B SaaS solo. Receipts, not hype. Going deep on RAMSReady + FineFlip.

[George Bloor @Beakus]

36 products in a year. Combined revenue: about £79 a month.

The best one makes £31 and took a weekend. The one I spent five months on makes nothing.

Going first because nobody posts this bit. What quietly earns for you, and what are you still embarrassed about?

[George Bloor @Beakus]

A "no press release" claim in my own market research was false. There was one. Found it on the check, corrected the report and the brief, and episode 1 stays held. "They never said X" is the easiest claim to write and the hardest to prove.

[George Bloor @Beakus]

21, 22 and 18 out of 40. That's how three market ideas scored on my rubric, so all three are dead and I'm staying on the bet I'd already made. The research didn't find a winner. It stopped me chasing three losers.

[George Bloor @Beakus]

84 cron jobs on my server now report in when they fail, instead of dying quietly. Then I ran a kill-Claude drill: took away the model my agents run on to see what broke. It passed. Boring infrastructure is what lets me trust the clever stuff.

[George Bloor @Beakus]

287 videos from 5 founder channels turned into one playbook overnight. The other 8 channels have been stuck behind YouTube's bot wall since 01:27 and the queue's still retrying. Summarising was the easy bit. Getting the transcripts is the job.

[George Bloor @Beakus]

18 of 18 videos have passed the fact-check gate since I added it to my YouTube pipeline. Reach at 3 days is 106 against a median of 548. The comments aren't about the facts, they're about the animation. The gate stays. The animation doesn't.

[George Bloor @Beakus]

The default path in my generated webhooks was /api/track/*. It 307s to the login page. The public one is /api/proxy/track/*. Fixed on one product, the others likely have the same bug. A 307 doesn't look like a failure until you go looking.

[George Bloor @Beakus]

250 visitors a day, 1% start the builder, £0. Weekend baseline on one of my SaaS tools. I wrote the fork down before the weekday numbers landed: converts, scale it. Doesn't, kill the cold consumer push and go B2B. Set the rule before you see the data.

[George Bloor @Beakus]

Ran the numbers on a calibration rig I'm building. Payback from saved calibration fees: 6.2 years. So it doesn't sell on cost savings, it sells as audit insurance or not at all. Better to find that in a spreadsheet than after the build.

[George Bloor @Beakus]

Free tier now returns the whole document, stamped SPECIMEN: 140 tiled text marks that survive printing. The flag sits in a database column, not the client, so there's no clean copy to leak. Verified on prod with a real free generation.

[George Bloor @Beakus]

Wired the YouTube ad for my RAMS builder straight through to the start page. Ad, builder, specimen document, £29 unlock now runs end to end. Google's policy check takes a business day, so Monday I find out if anyone actually walks the whole thing.

[George Bloor @Beakus]

My planning alerts were matching keywords by substring. Searching for "bar" hit BARNETBY. One council went from 57 matches to 5 once I made it match whole words. All 52 of those would have been a wasted email to someone with no job for me.

[George Bloor @Beakus]

Some planning portals ignore their own date filter. You ask for 7 days, you get everything. Looked fine because the count was plausible. Now I check the dates that come back against the ones I asked for, and bin the source if they don't match.

[George Bloor @Beakus]

Ran the payback numbers on a calibration rig I want to build. Best case 3.5 years on saved calibration fees. That isn't a product, it's a hobby. It only sells as audit insurance, or not at all. Better to find that in a spreadsheet than after spending £10.6k.

[George Bloor @Beakus]

Took a planning registry from 337 authorities to 407 last night. 57 of the 70 new ones needed no new code, they just dropped into existing adapters. The other 13 ate the whole evening. That ratio is why you build for variation early.

[George Bloor @Beakus]

Confirmed the Pfeiffer TPG serial protocol is fully documented: 39 mnemonics, ENQ/ACK handshake, per-reading status codes. That's the first driver I'll write, because a gauge that tells you the reading is bad is worth more than one that just hands you a number.

[George Bloor @Beakus]

Ran the numbers on a calibration rig I'm building. Payback on saved calibration fees: 6.2 years for the cheap build, 3.5 for the good one. That kills the pitch I had. It doesn't sell on saving money. It sells as audit insurance or it doesn't sell at all.

[George Bloor @Beakus]

@sl1ma4 Ask it for something small first. I got working code out of an agent in an afternoon, then spent three weeks on the boring 10% it got wrong. GTA 6 is about ten million of those. What's the biggest thing you've actually shipped from one?

[George Bloor @Beakus]

@DanKornas The hard part is not blocking the dangerous call, it is that approving fifty a day turns into rubber stamping by day three. Does yours learn what I've already approved, or is every review from cold?

[George Bloor @Beakus]

@DanielPriestley The trap I keep falling into is that the buyer only cares about the outcome, so anything upstream of that sounds like noise to them. How do you position without it turning into vague consultant-speak nobody can buy?

[George Bloor @Beakus]

@DanKornas The trail matters more than the agents. Most of my failures were unrepeatable because I could not see what an agent decided or why, only the diff after. Does it store the reasoning too, or just the job state?