DK
Back to Work
AI SaaS · Startup·Jan 2017 – Mar 2020

PixelAI.in

Co-founder

10K+Active users
40Countries
91%GPU cost reduction
ProfitableBootstrapped, no funding

Overview

PixelAI.in was an AI-powered image enhancement and restoration SaaS that I co-founded and built from zero. The platform offered photographers, e-commerce sellers, and designers tools like super-resolution, background removal, noise reduction, and colour grading — powered by custom deep learning models running on GPU inference infrastructure. We grew to 10,000+ active users across 40 countries and reached profitability without external funding.

The Challenge

The hard part was not the machine learning — it was serving GPU inference economically at sub-3-second latency for consumer users who expected instant results, while keeping infrastructure costs low enough to stay profitable on a $9/month subscription. Cloud GPU instances cost $2–5 per hour. A naive approach would make every free-tier request cost us more than the user paid in a month.

What I Built

01

Trained and fine-tuned ESRGAN (super-resolution), U-Net (background removal), and DnCNN (denoising) models on curated datasets — optimised models to TensorRT for GPU deployment, reducing inference time by 40% vs PyTorch baseline.

02

Engineered a GPU instance pool scheduler: auto-scales EC2 GPU instances (g4dn.xlarge) based on queue depth, pre-warms instances 2 minutes before predicted traffic spikes (from historical patterns), and hibernates idle instances — keeping GPU cost at $180/month vs $2,000+ with always-on approach.

03

Built the full-stack SaaS platform: Flask API backend, React frontend with real-time job progress (WebSocket), Stripe subscription billing with usage-based overage, team workspaces, API access tier with SDK, and a Figma plugin.

04

Designed the job queue architecture: image upload → S3 → SQS → GPU worker → result S3 → webhook/polling — fully async with retry logic, dead-letter queue, and cost-per-job tracking per customer.

05

Grew the product from 0 to 10,000 users through a SEO-first content strategy (long-tail "how to remove background from image" keywords), Product Hunt launches (#3 Product of the Day), and a free tier limited to 5 images/day.

Tech Stack

AI / ML

TensorFlowPyTorchTensorRTESRGANU-NetDnCNN

Backend

PythonFlaskCeleryPostgreSQLRedisSQS

Frontend

ReactJavaScriptWebSocketsStripe.js

Infrastructure

AWS EC2 (GPU)S3SQSCloudFrontRoute 53Docker
Next ProjectThe Branding Club