Skip to content
← All projects

Catalogue-to-Offer Agent for a Uniform Manufacturer

In ProductionJuly 1, 20262026

A hotel asks for uniforms for reception, waiters and housekeeping in its own colours. Before anyone orders, the manufacturer needs to show how those choices work together. This agent takes the proposal through selection, previews, revisions and a finished document in Telegram. Each decision stays connected to the catalogue, the agreed brief and the version a person actually reviewed.

A brief becomes a plan before it becomes an image

The Hermes conversation agent turns the request into roles, article codes, colours, page format, casting and logo placement. Catalogue, image and PDF operations run through separate MCP tools that check their own preconditions. The plan is saved with a content hash: changing a garment or colour requires a new plan. A confident sentence in chat cannot silently overwrite the agreed choices.

Choose the main image first
After the plan is confirmed, three lighting previews let the person check faces, clothing and framing. The chosen main image is copied unchanged into the proposal and verified by hash. A different pose or missing full-length view starts a new plan and preview choice.
Keep spending separate from acceptance
Each new paid image action needs confirmation. Permission to generate is separate from approval of the finished sheet. Jobs retain their status and recognise repeated identical commands, so an interrupted run can be checked before another is launched.

A catalogue article and its photograph travel together

Each selected garment carries its article code, reference photograph and working colour into the next stage. For a new colour, the system looks for a photograph of the same model in the closest available colour and records the source; a compromise is disclosed. A colour change should retain the article, while a replacement names the old and new articles. For a pair, the two outfits are tracked separately.

New products can enter from Telegram
The manufacturer sends photos, a name and an article code, checks the proposed card and confirms it. Existing articles are checked to avoid duplicates. A listed product with no usable photograph needs a reference before it can be visualised, including a specific shoe model.
References guide the result
A real photograph gives the model a visible cut, collar, pocket and trim to follow. The generated image still needs comparison with that source. It helps discuss a design; physical fabric samples and fittings establish shade, composition, texture and fit.

The page is assembled from saved images

The image model produces individual photographic scenes. Code places them into a fixed template and adds captions, palette swatches and the header logo. Templates define the required cells for an individual, a pair or a reception proposal. A sheet can be rebuilt from its saved images without generating them again; a new format that needs an additional scene still requires that scene to be created.

Shared references across the scenes
Supporting scenes receive the catalogue photos, selected main image and casting profile as common inputs. Before assembly, the recipe is checked against the fixed plan. That keeps the instructions consistent; faces, garment details and lettering are still inspected in each generated image.
A generated outfit view used as one component of a proposal sheet. It shows the front and back for discussion and comparison with the catalogue.

The header logo and garment branding have separate rules

A client's website can supply a proposed palette and a logo for the sheet header. The person checks the colours, and the compositor places the original logo file with its proportions preserved. Branding on clothing requires a separate file and an explicit placement, such as the right sleeve of each shirt. That mark is generated with the fabric, light and folds, so its letters, shape and position need visual review. Header-only branding does not implicitly authorise logos on the garments.

Quality checks examine details, proportions and uncertainty

A vision model locates a requested collar, button or pocket and returns a bounding box. Code turns those coordinates into a crop from the saved image. A separate check asks whether the requested detail is visible, whether the crop contains more than background and whether it belongs to the same outfit. A missing required detail or failed mandatory check prevents the sheet from being completed and accepted.

Use the catalogue to check the main image too
For supported garments, checks compare proportions across frames. A usable catalogue photograph helps distinguish an incorrect main image from an incorrect supporting view. For a long apron, for example, its length and pocket position matter more than the overall impression of the scene.
An unreliable measurement can stay unknown
Folds, perspective and obscured edges can make a measurement unusable. The result can be recorded as unknown, and automatic shape correction is disabled where reliability is unconfirmed. Colour diagnostics compare groups of similar shades with the requested palette; they provide evidence for review rather than automatically accepting or rejecting a sheet.
A detail crop makes the neckline and lapel easier to inspect. The reviewer checks the visible image against the catalogue and the request.

A specific edit creates a new candidate

Moving or enlarging an existing crop changes the framing without calling the image model. Replacing a colour, button or garment is a generative edit: it records the source file and expected hash, creates a child version and reuses the saved neighbouring cells. The parent stays available for comparison. Cropping can reveal only what is already in the source; missing shoes or a new pose need a new main image and another selection.

Check the entire edited image
In a documented experiment, changing three gold snaps to black also changed pixels along the garment outline and side tie. A difference map made those changes visible. Saving the other cells protects the rest of the sheet, but the edited candidate still needs inspection for unintended changes to seams, shape and texture.

Approval belongs to an exact version and a person

When a person accepts a sheet, the system copies the sheet, its component images and original logo into an immutable revision. A file manifest records sizes and hashes and is checked when read; an incomplete or damaged revision cannot feed the lookbook. The record identifies the person who actually approved it, along with the plan, reference hashes, prompt version, generation settings and template snapshot.

Save the result as well as the recipe
A saved seed does not guarantee that an external image provider will reproduce the same pixels. Keeping the actual files allows the document to be rebuilt from the agreed images. Hashes establish file integrity; visual quality and acceptance remain human decisions.

A review image, a print sheet and a lookbook

The workflow produces distinct artifacts for discussion, printing and the whole project. The main lookbook uses saved, approved proposals without generating fresh images. Before assembly, the system checks the accepted revision, file hashes and project membership. Page order and the resulting PDF are then reviewed as a document.

PNG for review
One proposal sheet is shared in chat for inspection and edits. A separate overview image lays out the lookbook pages together so their order can be checked.
A4 PDF for one proposal
The print export places a single approved sheet into an A4 PDF. It is a separate output from the project lookbook.
Multi-page PDF for the project
The main lookbook contains a cover and a page for each approved proposal, using the studio's narrow vertical format. A pair stays together on one page. This gives the manufacturer a document built from the exact versions agreed during review.

What it proves

A production workflow connecting catalogue articles, approved plans, visual checks and saved revisions, with a person choosing the images and accepting the exact files used in the final PDF.

Category

AI Agents

Built with

Nano Banana Pro · Gemini API · MCP Server · Hermes · Telegram Agent · Pillow Compositing · Supabase · Docker

Need something similar?

The cheapest way in is two weeks. The first days work out which task would pay for itself in your processes; the rest builds that agent on your own data and measures it.