Core Engine Testing App Vibe Coded
Product Thinking AI System Design Built with Claude Code Full Stack

AI Content Studio

Hyperlocalized creatives for automobile dealers

Brands push the same creative to 500 plus dealers. Every dealer has a different city, buyer profile, and campaign need. This app tests the engine that generates dealer-specific social media creatives so a Diwali EMI post for a Pune dealer feels nothing like one for Ludhiana.

Hosted on Railway

Pipeline run completing Layer 5 image generation

Pipeline run completing Layer 5 image generation

Brand creatives don't localise themselves

In automobile retail, marketing flows top-down from brand to dealer. What arrives is polished, on-brand, and completely context-free.

Brand Side

Central teams produce high-quality campaign creatives for the whole network. One Diwali post. One EMI banner. Designed to work everywhere, which means optimised for nowhere.

Dealer Side

A dealer in Nagpur running a Thar EMI offer speaks to a different buyer than one in Gurugram promoting a test drive. Generic brand creatives do not speak to the local context.

The Gap

No hyperlocalization. No occasion-specific framing. No variety across 500 plus dealers. The creative looks like it belongs to the brand, not to the dealership.

The Hypothesis

If a dealer can describe their context in structured inputs, a generative AI pipeline can produce a creative brief and image specific to them, at scale, without a design team.

Why this needed AI

CONTEXT

Every dealer input is a unique combination

City, occasion, vehicle, buyer intent. No two dealers are the same.

Contextual reasoning

BRAND CONSISTENCY

All creatives stay within brand guidelines

Dealer-specific does not mean off-brand. The engine works within guardrails set by the brand.

Brand-safe output

REASONING FIRST

Understanding context before generating

The model reasons about the dealer's situation before producing anything.

Thinking enabled

HUMAN IN THE LOOP

Dealer makes the final call

AI handles creative range, the dealer selects what fits their market.

Human oversight

5 layers, 2 human touchpoints

The pipeline separates reasoning, creation, and judgment into distinct steps. Each layer has a clear input, a clear output, and a clear owner.

7-step structured input form

7-step structured input form

Dealer Input

7 structured fields. Purpose, occasion, vehicle, localization level, human presence, reference asset, dealer identity. No free text.

YOU

Reasoning Note

Before any creative output, the model reasons about buyer mindset, occasion context, and what the creative needs to accomplish.

Gemini 2.5 Flash

Story Line Generation

3 story lines with different tonal anchors using the reasoning note as grounding.

Gemini 2.5 Flash

Story Line Selection

Dealer picks one direction. Second and final human touchpoint.

YOU

Prompt Engineering

Translates selected story line into a structured image generation prompt. 1:1 format for social media.

Gemini 2.5 Flash

Image Generation

Final creative rendered as a downloadable image. Stored in Postgres, not the filesystem.

Gemini 2.0 Flash

This is a testing instrument, not a demo

Every design choice, run history, thinking traces, annotation panel, per-layer inspection, exists because the real goal is to test and tune the engine.

PRODUCT INSIGHT

Structured input is a product lever

Constraining what the user can say improves what the model generates. Input structure matters as much as prompt engineering.

MODEL BEHAVIOUR

Reasoning before generation matters

Layer 2a before Layer 2b produces noticeably more grounded outputs than going straight to generation. Thinking budget size affects this.

UX DECISION

The right human-in-the-loop moment

Story line selection, not image approval, is the right point to involve the human. Narrow enough to be useful, open enough that choice still matters.

NEXT VERSION

What v2 would look like

Reference asset upload, multi-format output (1:1, 4:5, 16:9), and a scoring layer that evaluates creative-to-brief alignment before showing the dealer.

Built with Claude Code

Full codebase, React frontend, Express backend, 5-layer pipeline, Railway deployment, built through Claude Code. Phase 9 complete.

Infrastructure

Stack

Frontend

React + Vite

Plus Jakarta Sans

Backend

Node + Express

Async pipeline

Reasoning

Gemini 2.5 Flash

Thinking enabled

Image Gen

Gemini 2.0 Flash

Inline image

Database

PostgreSQL

Railway + SQLite local

Hosting

Railway

Frontend + Backend

Version Control

GitHub

main + dev