.svg){kind=icon}
DeepSift
AI Image Detector
DeepSift tells you in seconds whether a photo is human or AI. Taqwah designed it all from scratch: identity, web app, landing page, and mobile, as one connected system.
The Project,
At a Glance
Ready-to-build product, covering research, UX, brand identity, and mobile design.
.png)
.png)
Someone on Our Team
Was Harassed With a Fake AI Image
We looked for a simple tool to prove the image was fake. Nothing worked well
- Existing detectors were built for researchers, not everyday people
- Most tools showed confusing accuracy scores with no clear answer
- Uploading a personal photo to unknown platforms felt unsafe
- The people most affected by fake AI images had the least access to reliable tools
%20(3).png)
%20(4).png)
Brand Identity Design
for DeepSift
Most AI detection tools look like lab software. DeepSift needed to feel like a friend, not a forensic system. A calm palette and clear typography make the product feel safe in the moments people need it most.
.png)
%20(1).png)
%20(2).png)
AI Detection,
Fast, Clear, and Safe
Recording had to feel effortless. The AI summary had to earn trust. And the path from "I recorded something" to " have a proposal ready" had to feel natural, not like extra work.
One Clear Verdict
Users needed a plain human or AI answer, not a percentage score they had to decode.
Built for Non-Experts
The interface had to work for a parent, a student, or a journalist. Not just engineers.
Safe to Upload
People upload sensitive images. Privacy had to feel obvious, not buried in a policy page.
Verdict First, Then Trust,
Then Everywhere
The user needed a clear answer before anything else. Every decision came after that: the calm brand identity, the safe upload flow, and the mobile experience. Answer the person, then build the trust around it, then take it everywhere they need it.
.png)
.png)
.png)
Mobile App Design for
Checking Images on the Go
Most suspicious images arrive on a phone. The mobile experience keeps the same one-tap upload, clear verdict, and safe flow, designed for checking images the moment they land in your inbox or DMs.
.png)
.png)
.png)
.png)
AI Image Detector
Web App UI Design
The core product. Upload a photo, get a verdict in seconds. Every screen removes friction: a drag-and-drop upload, a live analysis state, and a result page that says human or AI in plain language before showing the technical detail.
.png)
.png)
.png)
The UX Decisions
Behind the Product
A verdict you can’t see the reasoning for is hard to trust. Showing the percentages and confidence turns a black-box answer into evidence a user can stand behind.
Detection is only useful if it’s defensible. Supporting the score with origin evidence lets people justify the result, not just read it.
Verification tools can feel intimidating. A vivid, modern identity makes a technical, slightly scary task approachable for everyday users.
Every extra step costs a check that doesn’t happen. Letting people drop any common file straight in removes the barrier between question and answer.
People need to verify an image wherever they encounter it. One consistent experience across devices keeps the trust signal and the brand identical everywhere.
From Personal Problem
to Public Tool
Typing felt slow. Recording felt easier, but recordings piled up unused. Nobody wanted to record, then transcribe, then edit, just to get one clean paragraph.
Start With the Real Scenario
A person receives a suspicious image and needs an answer fast. Every flow began here.
Simplify the Verdict
We replaced accuracy percentages with a plain human or AI result. Technical detail lives one tap deeper.
Build the Brand for Trust
Warm colors, clear language, no lab aesthetic. The identity signals safety, not surveillance.
Landing Page That Moves
Doubt to Answer
The landing page had one job: get someone who just received a suspicious image to upload it and check. Clear messaging, real examples, and a conversion flow built to turn worry into action in one click.
.png)
.png)
AI Image Detector Ready to Build
0.98 AUC
DeepSift was benchmarked against a messy real-world mix, AI renders, edited photos, and untouched shots, not just clean lab samples. It separates generated from human images at a 0.98 AUC, which means the verdict holds up on the kind of images people actually paste in.
12+ Generators
New image models ship every month, and a detector trained on one is blind to the rest. DeepSift was tuned across Midjourney, DALL E, Stable Diffusion, Firefly, and others, so a fresh fake from an unfamiliar tool still gets flagged instead of slipping through.
Pixel Level
A score alone doesn't convince anyone. DeepSift highlights the regions that triggered the call, the over smooth skin, the warped text, the tell tale artifacts, so a moderator or editor can see why an image was flagged and defend the decision.
Batch Ready
One image or ten thousand, the same flow. DeepSift was designed to run drag and drop checks for a single user and bulk screening for a platform moderating a feed, so it fits a curious visitor and a content team on the same engine.
.avif)
.svg)
%201.svg)
