Danish Khan
Full Stack Developer
Danish Khan
Full Stack Developer
All ProjectsAiposters - Image Generation AI
Brand accurate image generator with social media scheduling
80%
Reduction in design-to-publish time
95%+
Brand consistency score across assets
4
Social platforms integrated
1K+
Image generations per day
Beta
Aiposters bridges the gap between AI image generation and brand consistency. Marketing teams can generate on-brand visuals using natural language prompts, then schedule and publish them directly to their social channels — all from one dashboard.
Most AI image generators produce generic outputs that don't match a brand's visual identity. Marketing teams waste hours manually editing AI outputs to match brand guidelines, then switching to separate tools for scheduling.
We fine-tuned image generation models on brand assets (logos, color palettes, typography) so every output is on-brand. A built-in content calendar with social API integrations (Instagram, Twitter, LinkedIn) handles scheduling and publishing.
I built the full stack — the FastAPI inference backend, LoRA fine-tuning pipeline, and the Next.js dashboard with the prompt editor and brand kit manager. I architected the brand asset preprocessing flow that converts uploaded logos and style references into fine-tune conditioning parameters, and wired up all four social OAuth integrations with the Redis scheduling queue.
Research
Studied how brand guidelines are structured — color tokens, typography rules, logo safe zones — and mapped these to generation model conditioning parameters.
Chose LoRA fine-tuning over prompt engineering alone because prompt-only approaches couldn't reliably maintain logo placement and brand color accuracy across varied content types.
Model Pipeline
Built a FastAPI backend that pre-processes brand assets into LoRA fine-tune datasets, then orchestrates Stable Diffusion inference with style conditioning.
Separated the fine-tune pipeline from the inference API — fine-tuning runs async on job completion, keeping inference latency independent of dataset size.
Dashboard
Next.js frontend with a prompt editor, brand kit manager, and content calendar. Image generation results stream in via SSE for a real-time feel.
Used SSE over WebSockets for streaming — simpler infrastructure, works through Vercel's edge network without needing a dedicated socket server.
Integrations
Wired up OAuth flows for Instagram, Twitter, LinkedIn, and Facebook. Scheduling queue runs on Redis with a worker that handles retry logic and rate limits.
Redis over a database queue — needed sub-second enqueue latency and built-in TTL for scheduled posts, which a polling-based DB queue couldn't match cleanly.
Fine-tuning on brand assets is highly sensitive to dataset quality. A small, clean dataset of 30 high-quality brand images consistently outperformed a noisy 200-image dump. Quality over quantity is real — especially for generation models.
THANK YOU
" First solve the problem.
Then write the code."
~ John Johnson