TS Plastering: a Telegram-to-website pipeline that publishes case studies from photos
A bespoke website plus a Telegram-to-website pipeline. The plasterer sends a few photos and a one-line description after each job; GPT-4 Vision reads the images, GPT-4 Turbo drafts the case study, and the website queues it for one-click publish. Drafting effort per case study drops from 30–60 minutes to 5 minutes of operator review. Cost per case study: lunch money.
Labels distinguish client facts, reported results, comparisons and delivered system capabilities.
The bottleneck
TS Plastering is a one-team plastering and rendering business. The work is good, the photos prove it, and case studies on the website would compound into a portfolio that wins jobs. The problem: nobody on the team has the time, the writing inclination, or the CMS familiarity to turn finished jobs into published case studies.
The standard route: log into a CMS, upload images, write a description, format the layout, never happens after a 9-hour day. The result was a website that looked thin even when the business was actually busy.
What I built
Meet the operator where he already works. He uses Telegram all day to send photos to the office. So the case-study pipeline starts in Telegram.
- On-site. The plasterer snaps a few photos of the finished job. Captions them with a one-line description ("Kitchen replaster, Fleetwood, skim coat finish on damp wall"). Sends them to the bot. Total time: about 30 seconds.
- AI processing. The Telegram bot triggers a Supabase Edge Function. GPT-4 Vision reads the photos to extract job scope, surface conditions, and finish quality. GPT-4 Turbo composes the case study: title with location and technique, full narrative, structured "work completed" (challenges → approach → result), and a materials list. Total AI time: around 60 seconds.
- Pending review. The draft lands in the website's admin queue with a "pending review" status. The operator (Dean or the business owner) reviews it in about 5 minutes, edits anything the AI got wrong, and publishes.
- SEO-aware. AI-generated content is keyword-conscious by design: location, finish, technique, and materials all appear naturally in titles and descriptions. Each case study compounds into the site's SEO surface.
Built bespoke on React + Supabase + Telegram, with OpenAI for the vision and language calls. Hosting runs on the Supabase + Netlify free tiers, total ongoing infrastructure cost: zero.
What changed
- Case-study publication time drops from "never happens" or 30–60 minutes per job to ~5 minutes of operator review.
- Cost per case study: £0.10–£0.30 in AI usage. Lunch money.
- SEO surface compounds with every job, each case study is a fresh, indexable page with location and technique keywords already present.
- Workflow respects the operator. The plasterer never opens a CMS. He uses Telegram. The system meets him there.
By the numbers
The operational economics of running a Telegram-to-website AI case-study pipeline for a single-team trade business.
Services used in this build
The capability and service pages this engagement delivered against. Use them to scope a similar build for your operation.
Custom AI Systems →
GPT-4 Vision reads on-site photos and extracts scope, materials, and finish quality.
AI Apps & Websites →
The bespoke trades website that the AI pipeline publishes case studies into.
Document Automation (use case) →
Photos → structured case study, written and published in ~5 minutes at £0.10–£0.30 per published item.
AI Workflow Automation →
Telegram-to-website pipeline running end-to-end with operator approval inside the loop.
Questions about this build
Because the plasterer is on the tools. He is not going to log into a CMS, fill in a form, upload images, and write a description after a 9-hour job. Telegram is where he is already messaging, taking photos and sending them takes 30 seconds. The bot picks up from there. The choice of channel is the whole point; a Make.com webhook or a manual upload form would have killed the workflow before it started.
GPT-4 Vision reads the uploaded photos to extract scope (kitchen / bathroom / lounge), surface conditions (cracks / damp / damage type), and finish quality. Combined with the plasterer's brief Telegram caption, GPT-4 Turbo drafts the full case study: title with location and technique keywords, narrative description, scope, materials used (e.g. British Gypsum MultiFinish), and the work approach. Everything lands in a "pending review" state on the website until the operator approves it.
Occasionally. That is why the pending-review workflow exists. The operator scans the AI draft, fixes any specifics (e.g. wrong material name), adds a personal note if relevant, and publishes. Total review time per case study: about 5 minutes. Without the AI, drafting a comparable case study from scratch is 30–60 minutes per job. With the AI, the bottleneck moves from drafting to review.
About £0.10 to £0.30 per case study in AI usage (GPT-4 Vision and GPT-4 Turbo combined). Hosting is free at this scale (Supabase free tier covers the database, storage, and Edge Functions; Netlify free tier covers the website). Total ongoing cost: lunch money. The bespoke build itself is the only investment.
Still have a question? Book an Operations Review, direct line to me, Dean.
Other industries I've built for
Different operations, same engineering discipline.
Want a build like this for your operation?
Every build starts with an Operations Review. Map your bottlenecks. Get a costed bottleneck map, whether we build or not.
Book an Operations Review