How the AI works

Five passes, one rule: only your facts.

Reading design work and writing an honest narrative about decisions is not possible without modern vision and language models. Here is exactly what they do, and what they are not allowed to do.

The pipeline

What happens in the minute after you press Draft.

Vision

Read the images

Each image is downscaled and sent to a vision model with its file name. It identifies what it shows: a wireframe, a final screen, a site plan, a sketch, a photo of a shelf.

Language

Understand the brief

Your notes, role, timeline, metrics, and earlier answers are read together. Casepanorama finds the problem, who you worked with, and what limited you.

Structure

Structure the story

The model returns one strict JSON object: problem, role, constraints, decisions each with a reason, outcome. It is validated with a schema; anything malformed is rejected.

Sequence

Sequence the visuals

Images are ordered as a narrative: context or before, research and sketches, iterations, the final work, details. Each gets a caption of what a reader should notice.

Honesty

Ask for missing facts

Where a number, a before state, or a reason is missing, it writes a question instead of a sentence. Your answers become facts for the next draft.

The facts guard

Numbers you did not type never appear.

After the model answers, a deterministic guard runs on every sentence. Any sentence containing a number, percentage, or amount that does not appear in your notes, fields, or answers is removed before you see the draft. The demo tells you how many sentences it removed.

The same pass enforces house style: no em dashes, and none of the empty marketing words on our banned list. The system prompt also forbids invented clients, awards, team sizes, dates, and outcomes.

Model-agnostic

Routed across providers, not married to one.

Casepanorama talks to models through one internal layer. Providers are tried in a configured order; if one is slow or down, the next answers, and the caller never knows which model ran. Images are translated into each provider's own format inside that layer. Agency plans are routed to higher-capacity models first.

If no model is reachable, a deterministic house writer takes over. It sorts your own sentences into the structure, pairs each decision with the problem it answers, orders images from visual cues, and turns every gap into a question. It is plainer than a model, and it never adds a fact.

The contract

The shape every draft must take.

Validated with zod on the server. Image order must be a permutation of your uploads; captions line up one to one.

{
  "title": "...",
  "summary": "...",
  "sections": {
    "problem": "...",
    "role": "...",
    "constraints": ["..."],
    "process_steps": [{ "step": "...", "decision": "...", "why": "..." }],
    "outcome": "..."
  },
  "image_order": [2, 1, 3, 0],
  "captions": ["...one per image..."],
  "missing_facts": ["One metric would help if you have it."]
}

Your data

What the models see, and what they keep.

Models see only the case study being drafted. Providers are used under terms that exclude training on API inputs, and Casepanorama never uses your uploads to train anything. The homepage demo keeps nothing: images and notes live only in the memory of one request.

Read our care standards · Your data