3
products shipped
sherifbutt.dev
· cardiff, uk
$ whoami
Hi, I'm Sherif. I build AI tools that ship.
Founder & technical operator in Cardiff. A decade of full-stack, three years deep on AI systems — backend, frontend, ML pipelines, infra, end-to-end. Currently shipping TrueStage (live, profitable) and building Loyal Receptionist (a 24/7 AI receptionist for UK trades) alongside Loyalleads. Open to investor conversations and select studio engagements.
● loyalreceptionist · pilot
truestage.co.uk · shipped · profitable
loyalleads · studio
10+
yrs full-stack
3
yrs on AI
1
studio
##now
updated september 2026company
Loyalleads — websites for businesses. the limited company the other two trade under.
selling
Loyal Receptionist — 24/7 AI secretary. calls, SMS, WhatsApp, bookings.
selling
TrueStage — AI property staging. for UK estate agents.
##what i'm building
click to expandshipping · paid
-
01› TrueStage — AI property staging for UK estate agents.paidEmpty rooms to buyer-ready in 24 hours. Intelligent pipeline trained on UK retail references — John Lewis, IKEA, Next. Watermarked for transparency. First 5 stagings free for new agents.
open-source
-
02› claude-image-tts-gen — Claude Code plugin for image + TTS generation.ossOne plugin, every provider: Google, OpenAI, OpenRouter, ElevenLabs, and any local OpenAI-compatible server (Kokoro-FastAPI recommended). Drops straight into Claude Code workflows.
-
03› voice-of-customer — Mines Reddit and HN for how people actually describe a problem.ossClaude Code plugin — MCP server plus skill. Pulls the verbatim complaints out of Reddit and Hacker News, so copy answers a real objection instead of an invented one. MIT.
in validation
-
04› Loyal Receptionist — 24/7 AI receptionist: calls, SMS, WhatsApp, bookings.pilotAnswers the 11pm boiler call, books it, and texts the customer when you are running late. Operator mode lets the owner run the business by voice. English and Cymraeg. Built into a Loyalleads site, or standalone.
archived · with receipts
-
05› SpeechBox — The Ollama experience for speech models.parkedLocal-first. One-click install. OpenAI-compatible API. TTS and STT models in a single binary. The gap LM Studio leaves open.
-
06› FlowScrape — Agentic scraping platform · 2 years R&D.parkedTaught me: go narrow. The general-purpose agent is a product trap; a vertical tool that does one scrape really well is what people pay for.
-
07› AI Personality Writer — Character engine with retrieval memory.parkedTaught me: "memory for LLMs" is a feature, not a product. Parked until someone brings a specific wedge to build around.
##writing
all posts & drafts →Aug 2026
The prompt I never wrote
None of the things I say when I want an image contain the word image. Getting an agent to notice that worked — and made everything downstream more expensive, harder to reproduce, and impossible to give up.
10 min
Jul 2026
I built a tool to find customer pain. Its best output was telling me I was wrong.
Reddit blocks AI crawlers, so I built an MCP plugin to get in. Then it spent three days proving me wrong — about my market, about my wedge, and finally about whether it should be a product at all.
11 min
Apr 2026
Using YOLO for CAPTCHA recognition — 17 phases from demo-ware to production
Quad-signal fusion, 7-layer motor control modelled on neuroscience papers, cost/accuracy tables. The naive approach works in the demo and breaks in the wild — this is the rebuild.
14 min
May 2026 · draft
The dark side of open source in the age of AI
AI collapsed the cost of replicating a SaaS from $500K to $500. That changes the open-source calculus — here's the math with Sentry / HashiCorp / Redis as the evidence.
~10 min
May 2026 · draft
Self-healing selectors: why V-Scrape rebuilds its own extraction logic
Cached LLM-derived extractors + drift detection. How to not pay for the same reasoning twice.
~7 min
Jun 2026 · draft
BYOK is a feature, not a cost centre
Why passing AI provider keys through to the user is the right default for power tools.
~5 min
Jun 2026 · draft
Benchmarking scrapers honestly — methodology behind ScrapeMe
What the 18 scenarios test, why 1,800 points, and how the 9-signal bot-detection score is computed.
~9 min
3 published · 5 drafts · updated monthly
browse the full index →
##support the work
if something here helped##contact
one url, one inboxThank you.
Your message is on its way.